Add inference_batch for sharing const-prop results across compilations

[AntonOresten]

Each compilation's interpreter starts with an empty local inference cache, where Julia keeps const-prop results for callees, which never reach the global code cache. A batch of compilations of one method with different constants (an autotuning sweep) re-derives them every time. cuTile worked around this by threading one cache through a sweep's interpreters via a package-local ScopedValue (JuliaGPU/cuTile.jl@6dcda23): 9.7 → 2.4 ms per config.

inference_batch(f) is that in GPUCompiler, bounded to where sharing is sound: interpreters constructed within the scope share a cache per task, cache_owner and world. Outside a batch nothing changes; get_interpreter is untouched, and back-ends with their own interpreter can call inference_cache(owner, world).

Toy measurement (24 const-propagated inlineable stages, a per-variant @noinline callee, 25 variants): inference 8.8 → 4.9 ms per variant. End-to-end compile(:llvm) is unchanged here, where LLVM dominates; the gain is where inference is the compile.

Needs JuliaGPU/CompilerCaching.jl#24 to be effective: without it the callee walk clears the shared cache in place after any compilation with a source-less callee (measured: no gain). Adds ScopedValues.jl as a dependency (also in #915).

Made with Fable

[AntonOresten]

Closing for now. Wasn't sold on the idea and it's very targetted