MilkdropPreset: add opt-in preprocessed-HLSL cache hooks (targeting 4.3)

[keeper-of-memes]

Targeting 4.3 (per the maintainer's guidance in the preset-switch performance discussion). Will rebase once the remaining 4.2 items (#970, #971, #1003) have merged.

Problem

The Milkdrop shader transpiler runs the hlslparser text preprocessor (macro expansion) over each preset shader before parsing it. On slower devices this step dominates the CPU cost of a preset load: ~100 ms of a ~177 ms transpile measured on an Apple TV 4K / A15 (release build), which is a guaranteed dropped-frames hitch wherever the load runs on the GL thread.

The preprocessor's output is a pure, deterministic function of its input text (verified with a golden-output harness: byte-identical output and identical downstream GLSL), so it can be cached across runs, and even precomputed at build time for a bundled preset pack.

Change

• New C API: projectm_preprocess_cache_hooks (get/put/user) plus projectm_set_preprocess_cache(). The struct is copied by value; NULL clears. projectM stores no cache itself, so persistence and policy are entirely the application's. With no hooks registered, behaviour is byte-identical to before.
• Mechanical extraction: the pure-text assembly of the preprocessor input moves verbatim from MilkdropShader::PreprocessPresetShader into the GL-free free function AssembleApplyPreprocessorInput() (new MilkdropShaderPreprocess.{cpp,hpp}), so external tooling (e.g. a build-time cache generator) can produce byte-identical preprocessor input without a MilkdropShader instance, GL context, textures, or PresetState.
• Keying: ComputePreprocessCacheKey() hashes the exact bytes handed to ApplyPreprocessor: two independent 64-bit FNV-1a passes plus the exact length (~128 bits effective), prefixed with a compile-time version salt. The salt is bumped whenever the assembly transform, static shader header, or hlslparser changes, so stale entries simply miss and fall back to the live preprocessor. Wrong output is impossible; the worst case is a cache miss.
• Integration: TranspileHLSLShader() consults the hooks before ApplyPreprocessor (a hit skips it entirely) and stores after a miss. Hook invocations happen on the thread that calls the projectM rendering functions, so no extra synchronisation is needed.

Measurements

With a build-time prepopulated cache for a bundled preset pack (tvOS/ANGLE, A15): preset-load transpile 177 → ~77 ms. The residual is parse/generate and the GL compile, not the preprocessor.

Notes for reviewers

• Zero behaviour change when unused; the only unconditional work is a null-pointer check.
• The RenderContext gains one pointer field to propagate the hooks from the instance to the shader code.
• The big-looking MilkdropShader.cpp diff is dominated by the verbatim code move into MilkdropShaderPreprocess.cpp.
• Builds clean on macOS (Release), unit tests pass.

[keeper-of-memes]

Update from production use: the cache is now deployed on Apple TV HD (A8) hardware as well, where it matters considerably more than on the A15 it was measured on — cold-cache preset loads on the A8 run into whole seconds, and the prepopulated cache absorbs the preprocessor share of that on first load. Also note this branch already includes the fix for computing the cache key when only the put hook is registered (previously a put-only registration would store under an empty key). Still targeting 4.3; will rebase and move the API into a dedicated header once #970 lands.