Class CacheState

CacheState manages domain-specific caches for the WebGPU runtime.

Currently contains:

• shapeCache: Caches TVM ShapeTuple objects keyed by dimension string.
  • Why: makeShapeTuple() is called on every tensor operation, crossing the JS→WASM FFI boundary each time. During LLM decode, the same shapes repeat every token (e.g. [1,32,128]), so caching avoids thousands of redundant FFI round-trips.
  • Invalidation: Cache entries may be evicted, but returned shape tuples hold independent references and remain valid for their caller's scope.

Future additions (follow-up PR):

• uniformCache: Caches GPU uniform buffers keyed by content hash.
  • Why: Many dispatches use identical scalar arguments (matrix dims, etc.). Reusing the buffer avoids createBuffer + writeBuffer overhead.
  • Invalidation: Must invalidate on any GPU buffer deallocation, since buffer pointers can be reused by the allocator, making cached entries that reference the old buffer stale.