Run every registered code action, not just the applied one

[meziantou]

Stacked on #1354, which fixes the first bug this change found. Review that one first; this PR is red without it.

Problem

Two gaps meant most code fix providers were never really exercised.

1. A provider configured without an expected fixed code was never invoked at all. ValidateAsync only calls VerifyFix when ExpectedFixedCode is not null, so .WithCodeFixProvider<T>() is a no-op otherwise. 355 of the 989 test methods that expect a diagnostic configure a provider that never runs:

test class	methods
DoNotUseBlockingCallInAsyncContextAnalyzer_AsyncContextTests	64
DoNotUseImplicitCultureSensitiveToStringAnalyzerTests	42 (2 fix assertions in 88 tests)
AvoidUnusedInternalTypesAnalyzerTests	36
UseIFormatProviderAnalyzerTests	20
UseStringComparerAnalyzerTests	18
NamedParameterAnalyzerTests	16

2. Only one action per diagnostic was ever run. A provider commonly registers several actions; VerifyFix applies actions[codeFixIndex ?? 0] and ignores the rest, so the siblings were unexercised even in tests that do assert a fixed code. And registering a fix does not run it — the delegate passed to CodeAction.Create, where the fix actually builds the new document, only runs when its operations are computed. That delegate is where the casts, the argument indexing and the rewriting live.

Change

• When no expected fixed code is declared, register the fixes for every reported diagnostic the provider declares as fixable.
• In both paths, compute the operations of every registered action, which runs each delegate, and let an exception fail the test.

The changes are still not applied when the test declares no expected fixed code — there is nothing to compare them against. Diagnostics are matched to their document by syntax tree, so a diagnostic from an API-reference document is not handed to a provider along with the wrong document. VerifyDiagnostics now returns the documents and diagnostics it already computed so nothing is recomputed. GetSortedDiagnostics had one caller and is inlined into it.

It found a real bug immediately

The first run surfaced an InvalidCastException in OptimizeStringBuilderUsageFixer.RemoveToString for sb.Insert(0, 10.ToString()) — the fixer reads Arguments[0], which is the index for Insert, not the value. Fixed in #1354, which this is stacked on.

Note the earlier version of this PR, which only called RegisterCodeFixesAsync, would not have caught it: the crash is in the action body.

One test had to change

Test_LongElseIfChainWhereEveryBranchJumps_AllElsesReported (added in #1352) now builds an analyzer-only ProjectBuilder. An else if chain is nested syntax, so its 1000 branches are a 1000-level deep tree, and the formatter Roslyn runs while a code action computes its changes recurses once per level and throws InsufficientExecutionStackException. That is a Roslyn limit on deeply nested syntax, not a defect in the fixer, and the test exists to guard the analyzer's complexity.

What this does and does not catch

Catches a provider that throws on a shape no fix assertion covers — unguarded casts, argument-index arithmetic, Single() on an empty sequence.

Does not catch a provider that produces wrong-but-compiling output; that still needs a real ShouldFixCodeWith. The MA0090 bug in #1351 was of that second kind and would not have been caught here.

Cost

Full roslyn5.9 suite: 59.9s on the base branch, 1m02s with this change — about +3s (5%).

All five versions pass: 3708 (4.8) / 3756 (4.14) / 3783 (5.0) / 3807 (5.6) / 3840 (5.9), zero failures.

Not in this PR

Two related harness gaps, left out to keep this reviewable:

• ValidateAsync's opening guards are dead code — DiagnosticAnalyzer and ExpectedDiagnosticResults are IList<> initialized to [], so is null is never true and a test that forgets .WithAnalyzer<T>() passes silently.
• ParseSourceCode drops a malformed [|…|] marker without warning, turning a positive test into a vacuous negative one.

A stricter follow-up is also possible: assert that a registered action actually changes the document, which would enforce the AGENTS.md rule against registering a fix that returns the document unchanged. That one will surface real failures, so it is worth doing separately.