Cut per frame, per verification and per scan work found in the review

[SimonCropp]

The perf items from the review in todo.md that could be changed and checked here.

• Queue projection grouped entries and detected label collisions pairwise, every frame and twice per change under the lock. Group keys are now worked out once per entry and collisions counted in a dictionary.
• ProcessCleanup built a string of every process's command line on each passing verification, even with logging off. It is now only built when logging is on.
• The selection summary built the whole selected text every frame to measure it. It is now counted from the spans.
• DiffPlex diffed the words of every modified line pair to fill sub-pieces nothing reads. Lines are now their own single word, so that pass is trivial.
• The Windows frame wait slept about 31ms rather than 16, and kept waking 60 times a second while hidden. It now waits for input or the next frame, and 100ms at a time while hidden.
• Long rows were laid out whole every frame, though nothing scrolls horizontally. The Windows canvas and the native payload now clip them to the window.
• raylib spun for the end of every frame. Its partial busy wait is off.
• InlineStaging.Clear read and parsed every staged patch on each verification. It now re-reads a directory only when its write time changes.
• The tray's accept retried a move eight times over three seconds on the UI thread even when the target was read-only or its directory missing. It now gives up at once for those.
• The tray's scan re-read every equal-size, differing pair every two seconds. A pair found different and unchanged since is skipped.
• PiperClient connected to the tray's port after the tray had exited, which costs two seconds where the SYN is dropped. It now checks the listener table first, and traces the failure as before. The SendOnly snapshot changes to say nothing was listening.

Not done: the attached viewer's five-a-second full listing (needs a protocol change), macOS repainting every frame (Swift, which I could not build here), and caching scaled images in the Windows head.

The native changes (CMake flag) mean this push runs build-native, which should open its own PR with rebuilt binaries.

Run locally on Windows, all passing: DiffEngine.Tests 703, DiffEngineViewer.Tests 359, DiffEngineTray.Tests 246, and a Release build of the solution.