[5.0] Drop [ThreadStatic] walker cache and allocate walkers on demand

[josefpihrt]

Summary

The hand-rolled [ThreadStatic] _cachedInstance + GetInstance()/Free() pattern was repeated on 23 types. Reuse safety rested entirely on Free manually nulling every field, guarded only by Debug.Assert (a no-op in release); four call sites had no try/finally; and none of the collection-owning walkers capped the buffer they retained per thread.

This PR deletes that cache. Walkers are constructed when needed. Reuse setters became constructor parameters or read-only properties, which removes the stale-state class of bugs. RefactoringFlagsCache is deleted. The defensive walker.ContainsContinueStatement = false; in UseForStatementInsteadOfWhileStatementAnalyzer is gone, since the field can no longer arrive dirty.

Both StringBuilderCache copies are left alone. They already have a size guard, and one is vendored Orang code.

A shared ObjectPool<T> was prototyped for the eight buffer-owning walkers, then dropped: those allocations sit next to semantic-model work, and the pool protocol (reset ordering, copy-safe dispose, async/ref struct) was more machinery than the savings justified. If a regression shows up, ContainsLocalOrParameterReferenceWalker is the first place to look — it has 10 call sites, four of them inside loops in InlineLocalVariableAnalyzer.

Two bugs the pooling was masking

• RCS1187 reported inconsistently. UseConstantInsteadOfFieldWalker.CanBeConvertedToConstant had no initializer, so a fresh walker started at false, making ShouldVisit false and suppressing the diagnostic, while a recycled walker started at the true that Free had reset it to and reported normally. Whether RCS1187 fired therefore depended on whether the thread had run the analyzer before. The property now defaults to true, matching what Free always restored. The visible effect is that a static readonly field is now consistently reported when the containing type has a static constructor that does not assign it — the same answer the analyzer already gave when there was no static constructor at all. Covered by a new test and a ChangeLog entry.
• ConvertWhileToForRefactoring double-freed a walker. It called Free(walker) and then returned from inside the try, so the finally freed the same instance again, handing one walker to two future callers. Dropping pooling there removes it.

Test plan

• Full solution builds clean (dotnet build src/Roslynator.sln --no-incremental), 0 warnings, 0 errors
• Analyzers.Tests: 2236 passed (on the pooling iteration); after dropping pooling, related suites passed (RCS1187, unused member/parameter, make member readonly, mark const, use async await, remove redundant async await)
• Refactorings.Tests (472), CodeFixes.Tests (110), CSharp.Tests (15), Formatting.Analyzers.Tests (573), CodeAnalysis.Analyzers.Tests (52), Core.Tests, CSharp.Workspaces.Tests (34) — all passed on the pooling iteration; RemoveAsyncAwait refactoring tests passed after dropping pooling
• No [ThreadStatic] _cachedInstance remains outside the two StringBuilderCache files

Made with Cursor