Roll the per-thread stats into one collapsed breakdown per section

[erikdarlingdata]

Fixes #4575. Part of #4511.

Why

The plan viewer's Properties panel emitted one indented "Thread N" row per
thread, inline, right under each summary row: rows, rows read, executions,
elapsed, CPU, logical reads, physical reads, scans, and read-ahead reads. A
DOP 4 hash match already produces about thirty of those rows, and DOP 8
produces hundreds, so the handful of summary numbers most people open this
panel for were buried in a scroll marathon.

Ports erikdarlingdata/PerformanceStudio@ff7f1d9d59851b2f7438bdade83d710e4528a411
(and its current dev form, which is unchanged since that commit for this
code).

What changes

• The summary rows are untouched.
• The per-thread numbers now live in one collapsed "Per-thread breakdown"
  sub-expander per affected section (Actual Statistics, Actual Timing,
  Actual I/O), with threads grouped under a small header per metric.
• A metric with no per-thread data is skipped, so a section only grows a
  breakdown when there is something in it.
• Rows and Executions list idle threads too (a thread sitting at zero while
  its siblings work is exactly what someone opens the breakdown to see);
  Rows Read, Elapsed, CPU, Logical/Physical Reads, Scans, and Read-Ahead
  Reads only list threads with a nonzero value.
• The breakdown header carries a skew suffix when there is skew. The share
  test mirrors PlanAnalyzer's Rule 8 (Parallel Skew) exactly — same
  coordinator-thread filter, same 1,000-rows-per-worker floor, same 0.80 (≤2
  workers) / 0.50 (>2 workers) share threshold — so the header never
  contradicts the Parallel Skew warning the same plan raises. There is also
  an idle-thread check Rule 8 does not make: a thread that returned nothing
  at all while its siblings did real work reads as skew on sight even when
  the busiest thread is under Rule 8's own share threshold.
• The shaping (which metrics to show, in what order, and the skew math)
  moves into a new pure ThreadBreakdown static class in
  PerformanceMonitor.PlanAnalysis. The shared PlanViewerControl
  code-behind's new AddPerThreadBreakdown only renders what that helper
  returns; the previous inline nine-repeat block is gone.
• Reuses the existing OrangeBrush (0xFFB347) already used elsewhere in this
  same panel for warnings, rather than adding a new brush.

Lite: gets this too (shared control) — PlanViewerControl lives in
PerformanceMonitor.Ui, and both the Darling Viewer and Lite reference that
project, so this one change reaches both hosts.

Not ported: none. This PR is the whole ff7f1d9 diff for this file.

Test plan

New file: Darling/Darling.Tests/ViewerThreadBreakdownTests.cs, 12 pins on
ThreadBreakdown — no-more-than-one-thread returns null, an all-zero metric
returns null, idle threads included/excluded per metric flag, header text
and thread count, the skew suffix format, the 0.50 threshold at 3+ workers
(just over and just under), the 0.80 threshold at 2 workers (60/40 vs
81/19), the idle-thread-always-reads-as-skew case, the 1,000-rows-per-worker
floor, metric ordering/skip, and unit carry-through.

RED on dev (pre-fix 130bc6a9e): compile-only, since ThreadBreakdown
doesn't exist there yet — 13 CS0103 errors, confirmed in a detached
worktree at that sha with the new test file copied in.

Runtime mutation: changed the >2-worker share threshold from 0.50 to
0.90 in ThreadBreakdown.cs. Build_carries_a_skew_suffix_at_three_or_more_workers_over_the_fifty_percent_share
went RED ([FAIL]); reverted, rebuilt, GREEN again.

Ran on this rig (in-process, Darling.Tests.dll with
Microsoft.WindowsDesktop.App stripped from its runtimeconfig.json):

-class Darling.Tests.ViewerThreadBreakdownTests
-class Darling.Tests.ShowPlanParserCondAndMultiplePlanTests
-class Darling.Tests.ActualPlanRequestTests
-class Darling.Tests.ActualPlanDispatchTests
-class Darling.Tests.ActualPlanResultParseTests
-class Darling.Tests.QueryModificationDetectorTests
-class Darling.Tests.ActualPlanCaptureLoopTests
-class Darling.Tests.ActualPlanGatingTests
-class Darling.Tests.ReproScriptBuilderHardeningTests
-class Darling.Tests.DarlingAnalysisPipelineTests
-class Darling.Tests.DarlingMcpPlanToolsSurfaceAndSqlTests
-class Darling.Tests.DarlingMcpPlanToolsLivePostgresTests
-class Darling.Tests.McpPlanAnalysisEnvelopeTests
-class Darling.Tests.SerialLoopStoreSizeSourceTests
-class Darling.Tests.TsqlConventionGuardTests
-class Darling.Tests.DocCommentHygieneTests

Total: 238, Errors: 0, Failed: 0, Skipped: 2, Not Run: 0

Also ran every Viewer* class discovered in Darling.Tests (160 classes,
1,655 tests): 86 failures, all pre-existing and all FileNotFoundException: PresentationFramework — WPF types this Mac can't load, unrelated to this
change (the same classes fail the same way with this branch's changes
reverted). No PlanSync* classes exist in this tree.

Builds (Release, -p:EnableWindowsTargeting=true), each 0 warnings / 0
errors: PerformanceMonitor.PlanAnalysis, PerformanceMonitor.Ui,
Darling.Tests, Lite/PerformanceMonitorLite.csproj, Lite.Tests,
Darling/PerformanceMonitor.Darling.Viewer.

Screenshot plan (for the Darling Viewer and Lite — same control):

1. Capture an actual plan with a parallel hash match or sort at DOP 4+
   (rows in the thousands per worker, so the skew math has something to
   grade). Open the node's Properties panel.
2. Actual Statistics, Actual Timing, and Actual I/O each show their summary
   rows exactly as before, no inline "Thread N" rows under any of them.
3. Each affected section shows one "Per-thread breakdown (N threads)"
   sub-expander, collapsed by default, right where the old inline rows used
   to start.
4. Expand a breakdown: metric names appear as small headers (Rows, Rows
   Read, Executions for Actual Statistics; Elapsed, CPU for Actual Timing;
   Logical Reads, Physical Reads, Scans, Read-Ahead Reads for Actual I/O),
   each with its own indented "Thread N: value" list underneath.
5. If the sample plan has real skew, the breakdown header for Actual
   Statistics shows an amber "(skewed: N max / M min)" suffix next to the
   thread count, in the same amber this panel already uses for warnings.
   If balanced, no suffix.
6. A metric with no per-thread data anywhere (e.g. Physical Reads on an
   all-in-memory plan) doesn't appear in the breakdown at all.

CHANGELOG

SECTION: Changed
ENTRY: - The plan viewer rolls per-thread stats into one collapsed breakdown ([#4587]) - For an actual parallel plan, the per-thread stats in each properties section collapse into one "Per-thread breakdown (N threads)" section, as in PerformanceStudio, with a skew note in its header when the work is unbalanced, instead of an inline row per thread per metric.
REF: [#4587]: #4587