Avoid allocating per-instance lazy factory delegates using function pointers

[tannergooding]

Every lazily-computed Cursor/Decl/Type member is wrapped in a ValueLazy<T> whose factory is a captured Func<T>. Because the factory closes over the owning instance (and, in a number of cases, additional constructor state), each wrapper eagerly allocates a delegate -- and sometimes a display class -- in the constructor, none of which is needed once the value has been computed.

This replaces ValueLazy<T> with ValueLazy<TSelf, T>, whose factory is a static delegate*<TSelf, T> passed the owning instance explicitly. The function pointer is a static, so nothing is allocated per instance; the owning instance is threaded through GetValue(self) at the call site. Each captured lambda becomes a static factory method taking self.

---

Measured on the TerraFX Windows SDK generation (DirectX/d3d12) with DOTNET_TieredCompilation=0 (the NativeAOT-representative full-opt proxy the shipped tool runs under), sampled GCAllocationTick:

	Before	After
Total sampled allocations	280.3 MB	179.6 MB
Per-instance Func<T> factory delegates	~71 MB (the top allocators)	0

Every Func_1[...] (String, Cursor, Decl, TranslationUnit, Type, IDeclContext, ...) is gone. The remaining Func_2/Func_3 allocations belong to LazyList, which is a separate sealed class and out of scope here.

All 3706 generator tests pass and the build is warning-clean. -- The bulk of the diff is mechanical (326 fields, 327 factories across 172 files); it was produced with a Roslyn rewriter driven by the semantic model, so cases like the static TranslationUnit.GetOrCreate vs. the instance TranslationUnit property are disambiguated correctly rather than by text matching.

Note

This PR body was drafted by Copilot.