webgpu: the kernels take the canonical ABI, so marshal_ is gone

[yhirose]

Stacked on #4 (base backend-layer-3); this PR's diff is the one commit on top. Once #4 lands, retarget this to master.

Summary

The last item on the backend-layer list. WebGPU's WGSL kernels predated the shared kernel ABI: every entry point read one 96-byte Params uniform, a family picked its operation by a number in it, and the bind group was fixed (A and B read, C written, D and E read). webgpu.h carried marshal_ to translate each kernel id's canonical params into that layout.

Now WebGPU realizes a launch the way Metal and CUDA do:

• View i is bound whole at binding i, in the order gpu_ops.h lists the kernel's views. The binding after them is a uniform holding the views' element offsets, then the params, field for field as in gpu_abi.h. The offsets travel in the uniform because a binding offset must be 256-byte aligned, which a view's offset is not.
• A family's operation (add, sub, ...) is the pipeline-overridable constant OP, set per kernel id from a kernel table (kernel_, the counterpart of Metal's and CUDA's kernel_name_), not a params field.
• Each pipeline takes an auto layout from what its entry point declares, so the WGSL declarations are the kernel's side of the ABI.
• One launch_ serves both the shared ops (through dispatch) and the backend's own ops. The N-D ops' shape metadata is an ordinary input view into the meta ring.

Removed: Params, kernel_op_, marshal_, encode_, operands_, and the hand-built layouts. webgpu.h is about 290 lines shorter. docs/backends.md no longer lists anything as not shared.

Also fixes a latent hazard. launch_ could flush between reserving a meta slot and the launch that reads it. After the flush, a later call in the new batch could be handed the same slot and overwrite it before the launch read it. Reserving a slot now flushes first when the uniform ring is also full.

Test plan

• WebGPU under Deno, --gpu and --auto: 117/117 cases, 8789 assertions.

• Per-family dispatch counts are identical to the previous build in both modes, compared against a build of backend-layer-3. No op started falling back to the CPU.

• Mutations, each caught by the suite:

  • a wrong OP in the kernel table
  • a kernel reading one view at another view's offset (caught by the non-zero-offset conformance test)
  • the meta view losing its offset
  • two meta runs swapped

• Native Metal: 117/117 cases, 11640 assertions with the device; check_qwen greedy tokens match. gpu_host reference build: 117/117. CUDA host-side trace: identical. None of these compile webgpu.h.

• Not yet run in a browser. Everything here ran on Deno (wgpu/naga), and so does CI. Two constructs have not been exercised on Chrome's Dawn/Tint:

  • the same binding number declared with different access modes in different entry points
  • override constants

  WGSL allows both, but a run in Chrome is still worth doing.