[ExecuTorch][WebGPU] Steel q4gsw prefill GEMM — test coverage

Pull Request resolved: #20731

**Add steel-GEMM coverage to the `et_vk.linear_q4gsw` golden sweep (stacked on the steel op diff).**

The op diff routes M>1 q4gsw prefill to the new steel GEMM on a >=256-invocation device (`K % 16 == 0`), falling back to shmem/register-tiled otherwise. The existing M>1 CONFIGS (`q_proj_4k`, `gate_proj_pf`, `down_proj_pf`, `shmem_edge`) already exercise steel on such a device via the shape-discovering native sweep; this adds one small config that isolates the steel branch specifically and documents the routing.

**Changes:**
- `test_quantized_linear.py` / `test_webgpu_native.cpp`: add the `steel` config (M=96, K=2048, N=256) — below the shmem thresholds (K<4096, N<2048) so pre-steel it was register-tiled, which uniquely pins the steel branch; M=96 exercises the partial 64-row tile (edge masking).
- Document that M>1 `K % 16 == 0` shapes prefer steel on a >=256-invocation device (lvp) and fall back on a <256 device (SwiftShader) — the same fp64 golden validates whichever kernel runs.

Co-authored-with: Claude Code.
ghstack-source-id: 401515169
@exported-using-ghexport

Differential Revision: [D110660966](https://our.internmc.facebook.com/intern/diff/D110660966/)

Author: JCNTH

backends/webgpu/test/ops/test_quantized_linear.py

@@ -59,7 +59,10 @@ class Q4gswConfig:
     # requires N % 8 == 0 (torchao pads N for the scale layout), so odd-N / N=1 are
     # not exportable -- bicol's has1 odd-N guard is defensive (mirrors coop4's
     # general-N robustness) and unreachable through this op.
-    # Prefill shapes routing to the shmem GEMM (K>=4096 or N>=2048); M=128.
+    # M>1 prefill: prefer the steel GEMM (K%16==0) on a >=256-invocation device
+    # (e.g. lvp); else shmem (K>=4096 or N>=2048) or register-tiled (SwiftShader
+    # caps at 128). Same fp64 golden regardless of which kernel runs.
+    Q4gswConfig("steel", 96, 2048, 256),  # steel-isolating (K<4096, N<2048)
     Q4gswConfig("gate_proj_pf", 128, 2048, 8192),  # gate/up prefill (shmem via N)
     Q4gswConfig("down_proj_pf", 128, 8192, 2048),  # down prefill (shmem via K)
     Q4gswConfig("shmem_edge", 130, 4096, 2056),  # partial 32-tile bounds

backends/webgpu/test/test_webgpu_native.cpp

@@ -275,6 +275,8 @@ const Q4gswConfig kQ4gswConfigs[] = {
     // scale over 64-256 K-groups). q4gsw requires N % 8 == 0, so odd-N is not
     // exportable; bicol's has1 odd-N guard is defensive (mirrors coop4
     // general-N robustness).
+    // M>1: steel GEMM on a >=256-invocation device (K%16==0), else shmem/tiled.
+    {"steel", 96, 2048, 256, 1e-4f, 1e-3f, true, false}, // steel-isolating
     {"gate_proj_pf", 128, 2048, 8192, 1e-4f, 1e-3f, true, false}, // shmem via N
     {"down_proj_pf", 128, 8192, 2048, 1e-3f, 1e-2f, true, false}, // shmem via K
     {"shmem_edge", 130, 4096, 2056, 1e-4f, 1e-3f, true, false}, // partial tiles