Perf Phase 2: cached-symbol comparisons instead of formatted-string comparisons

[credfeto]

Part of #455.

Many sites format a symbol to a string per node purely to compare against a constant. Resolve the constant to an INamedTypeSymbol once at RegisterCompilationStartAction and compare with SymbolEqualityComparer instead (zero allocation, and gives a free bail-out gate when the type isn't referenced by the compilation). This is the pattern InternalsVisibleToDiagnosticsAnalyzer and ProhibitedSubstituteForUsageInTestBaseDiagnosticsAnalyzer already use correctly.

• Helpers/TestDetection.cs:44 — IsDerivedFromTestBase string-compares ToFullyQualifiedName() per base class per node. Convert to a cached-TestBase-symbol API (accept the compilation / cached symbol) and walk with SymbolEqualityComparer. Callers: TestClassAnalysis, TestClassFieldsAnalysis, TestClassPropertyAnalysis, ProhibitedSubstituteForUsageInTestBase.
• ArgumentExceptionAnalysisDiagnosticsAnalyzer.cs:129 — cache the 3 exception type symbols
• ClassVisibilityDiagnosticsAnalyzer.cs:138,179 — cache the MockBase<T> symbol
• DebuggerDisplayAnalysisDiagnosticsAnalyzer.cs:146 — cache the DebuggerDisplayAttribute symbol (also gives a compilation-start bail-out)
• SuppressMessageDiagnosticsAnalyzer.cs:90 — cache System.ReadOnlySpan<T> symbol
• ProhibitedClassesDiagnosticsAnalyzer.cs:91,217 — compare symbol identity against the already-cached type dictionary instead of re-formatting ToFullyQualifiedName() (currently done twice per node)
• ProhibitedMethodInvocationsDiagnosticsAnalyzer.cs:165,307,348,358 / ForceMethodParametersInvocationsDiagnosticsAnalyzer.cs:149 / ProhibitedMethodWithStrictParametersInvocationDiagnosticsAnalyzer.cs:120 — these rebuild a qualified-name string per node as the KEY into an already-cached dictionary. Re-key the caches by INamedTypeSymbol (SymbolEqualityComparer) so no string is ever formatted on the lookup path.
• ConstructorGenericParameterTypeDiagnosticsAnalyser.cs:147,311 — same treatment

Verification: full test suite green, benchmarks show reduced allocations for TestDetection-dependent and dictionary-lookup analyzers, no change in the set of diagnostics produced.