perf(recompiler): memoize SRT evaluation in a flat table and inline IrValue's hot accessors

[oneandonlydean]

What

Every draw stage's capture walks the shader's SRT plan with Detail::Evaluator. Its per-value memo was a std::unordered_map<IrValue*, std::uint64_t>: one node allocation per evaluated value. Every step of the walk also made out-of-line calls to IrValue::Opcode, IsIdentity, Resolve and Argument. In a profile of my integration branch at PPSA21564's title, the walk took about 17 us of every capture, about 1.25 s of the graphics queue worker's 10 s.

• The memo is now Detail::EvaluatedValues, an open-addressing table keyed by the value pointer (linear probing, power-of-two size starting at 64, grown at half load). Insert keeps the first value for a key, as emplace did, and failed evaluations are still not memoized.
• IrValue::Opcode, Type, HasImmediate, IsIdentity, ArgumentCount, Argument and Resolve move to the header as inline definitions, unchanged, including their throws.
• <unordered_map> is no longer used by SrtEvaluator.hpp and is dropped.

Evaluation order, results and failures are unchanged. In the same profile the capture walk drops to about 10 us (title capture total about 1.25 s to 0.7 s per 10 s).

Test: verifyEvaluatedValues in agc_shader_memory_tests inserts 1000 keys (several grows), finds each with its value (including a stored 0), checks that a second insert does not replace the first, and that empty tables and absent keys find nothing. The existing SRT walker tests cover the evaluator itself.

Relation to #218 (slick-daddy, "cut hot-path overhead in recompile cache lookups"): the two are complementary. #218 changes the evaluator's cycle stack (_visiting, a vector to an unordered_set) and the cache key, snapshot hash and memo lookups in Recompiler.cpp. This PR changes the evaluator's value memo (_cache) and the IrValue accessors. They touch adjacent lines in SrtEvaluator.hpp only: the member pair and the include block. Whichever lands second resolves to EvaluatedValues _cache; beside #218's std::unordered_set<IrValue*> _visiting;, with <unordered_set> kept. SrtEvaluator.cpp merges cleanly.

Shader disk cache: on main the cache's SourceVersion is a hash of the recompiler and AGC driver sources, so this change, like any edit there, gives a new version and cached shaders are recompiled once. The disk key format and the cached data are unchanged.

Tested

• Linux (CachyOS, GCC 16), RTX 4090: ctest 125/125 pass, including the new verifyEvaluatedValues in agc_shader_memory_tests.
• PPSA21564: on my integration branch through repeated passes of my visual regression check (intro, title, menu and a level). The timings above come from my profiling runs there.

Checklist

• Based on current main; no other open PR implements the same functions
• One topic per PR; follow-ups go in a new PR
• No comments in code except technical debt
• Unimplemented paths throw (NotImplemented_nid_no_patch); silent stubs are listed in TechnicalDebt
• No notes, investigation .md files or images added to the repository (attach them to this PR)
• New third-party code is a submodule built from source (none added)
• Generic behaviour, not specific to one title
• Depends on: none
• AI-assisted: yes

[github-actions]

Other open pull requests touch the same code. Merging one will break the other, or both implement the same thing. If this one builds on them, list them in Depends on.

Pull request	Conflicting files (lines on main)	Same exports	Same new files
#218	core/shader/recompiler/Optimization/include/Optimization/SrtWalker/SrtEvaluator.hpp L9, L36-37