fix: load Gemma-4 vision tower unquantized under compressed-tensors

[Smallfu666]

What

Addresses the Marlin shape-selection failure reported in #28018.

google/gemma-4-31B-it-qat-w4a16-ct ignores every vision-tower linear, but its clip
wrapper makes those ignore entries end in .linear. SGLang keeps that suffix on the
unfused clippable linears; the fused qkv_proj / gate_up_proj do not, so those two per
block miss the match and get quantized. gate_up_proj is 2 * 4304 = 8608, not divisible
by Marlin's min_thread_n = 64, so load dies in gptq_marlin_repack.

So drop the compressed-tensors config for the vision tower, with a warning_once. Other
quant methods are untouched: the NVFP4 checkpoint already excludes model.vision_tower*
by glob, and I have nothing justifying de-quantizing online fp8/awq/gptq.

Second point: CompressedTensorsWNA16 and GPTQMarlinLinearKernel both dropped the
return value of check_marlin_supports_shape, which never raises, so bad shapes hit the
kernel instead of the actionable min_thread_n = 64 error. Both now use
verify_marlin_supports_shape, like awq_marlin. I left the repack kernel alone.

Tests

CPU unit tests, each confirmed to fail when its own fix alone is reverted: one builds the
vision-tower modules against a two-layer replica of the checkpoint config and asserts
UnquantizedLinearMethod; two assert n=8608 raises ValueError on each Marlin path.

End to end on one GB10, same command both sides: main dies in gptq_marlin_repack
(size_n = 8608 vs tile_n_size = 64); patched loads the checkpoint and answers a
request.

Output quality not asserted; TP>1 and image input untested.

Happy to normalize .linear generically in compressed-tensors instead, if you prefer.

---

CI States

Latest PR Test (Base): ❌ Run #31200478757
Latest PR Test (Extra): ❌ Run #31200478387

[Smallfu666]

Superseded by #35792, which narrows the fix to the fused-shard .linear ignore mismatch and adds end-to-end load validation.