[Impeller] Compute dispatch contract cleanup: literal workgroup sizes, device limits, persistent bindings #193868

Part of #188474.

#188601 moved compute dispatch to 3D workgroup counts, but several backend differences remain in the dispatch contract that the Flutter GPU compute API will expose.

Changes

• Remove the Vulkan specialization constant that overrides constant_id = 0 with maxComputeWorkGroupSize[0] on every compute pipeline. It silently clobbers any user constant with that id. Convert prefix_sum_test.comp to a literal local_size.
• Validate the workgroup size against device limits when a compute pipeline is created, on Metal (the pipeline's maxTotalThreadsPerThreadgroup and the device's per-dimension maximum) and on Vulkan (maxComputeWorkGroupSize and maxComputeWorkGroupInvocations), and fail with a clear error.
• Expose the limits on Capabilities: max workgroup invocations, max workgroup size, max workgroup count, max shared memory and the minimum storage-buffer offset alignment.
• A dispatch with a zero count is a no-op on every backend instead of an error.
• Bindings persist across dispatches within a pass on every backend. Vulkan currently resets the pipeline and bindings after each dispatch.
• Fix the Metal compute pass pushing a debug group whose pop only happens in debug builds.

Testing

compute_unittests on Metal and Vulkan: a literal-size prefix sum, an oversized workgroup rejected at pipeline creation, a zero-count dispatch, and two dispatches sharing one set of bindings.

Dependencies

None.