Part of #188474.

ComputePass::Compute(ISize) is 2D and means different things per backend. Vulkan forces every shader's local size to the device maximum with a specialization constant and treats the argument as a workgroup count, Metal treats it as a total invocation count and divides by maxTotalThreadsPerThreadgroup. Both ignore the shader's declared local_size, which makes shared memory and 3D dispatch impossible.

Changes

• Replace Compute(ISize) with a 3D workgroup-count dispatch, for example Dispatch(ISize3 workgroup_count).
• Carry the reflected local_size on the compute pipeline and pass it to Metal's dispatchThreadgroups as threadsPerThreadgroup. Remove the Metal halving heuristic.
• Capture local_size in the reflection that feeds both the bundle and the generated .comp.h headers used by the HAL unit tests.
• Remove the Vulkan specialization-constant override and port the Vulkan path to the new signature so the tree compiles. Full Vulkan validation is tracked separately.
• Validate that local_size.x*y*z is within the pipeline maximum at creation.

Testing

Update compute_unittests to dispatch known workgroup counts and assert exact output element counts on Metal.

Dependencies

Pairs with #188477 to thread local_size from the bundle.