[perf] Check return literals before override guard in two return-type rules

[TomasVotruba]

NumericReturnTypeFromStrictScalarReturnsRector and BoolReturnTypeFromBooleanConstReturnsRector subscribe to ClassMethod. Both ran ClassMethodReturnTypeOverrideGuard::shouldSkipClassMethod() - parent/override reflection - before checking whether the returns are actually int/float (or true/false) literals. A method whose single return is not such a literal (the common case) paid for the override lookup only to then fail the literal check and return null.

The node-only literal check now runs first, so a method returning a non-literal expression bails before the reflection. The check reads only AST (Int_/Float_/UnaryMinus, ConstFetch + true/false), so the outcome is unchanged - the fail path returned null either way.

Behavior unchanged - all 8 fixtures across both rules pass.

Bench included (bench/return_type_literal_filter_bench.php), driving refactor() over 151 methods that return a non-literal expression. CPU per call, two runs each:

rule	before us/call	after us/call
NumericReturnTypeFromStrictScalarReturns	~31	~12.6
BoolReturnTypeFromBooleanConstReturns	~28	~11.7

~2.5x faster on that hot path. The remaining cost is findReturnsScoped(), which runs regardless. Run php bench/return_type_literal_filter_bench.php on each branch to reproduce.

Siblings checked and left as-is: StringReturnTypeFromStrictScalarReturnsRector already runs its literal check before the guard; NumericReturnTypeFromStrictReturnsRector has no node-only filter for the common case (its work is type inference).