[XPU] Add fused GemmaRMSNorm path for eager execution

[ccrhx4]

Purpose

GemmaRMSNorm on XPU fell back to separated elementwise ops under enforce-eager, causing a large prefill latency gap versus torch.compile (which Inductor-fuses the same ops).

• Add GemmaRMSNorm.forward_xpu that dispatches to the fused vllm-xpu-kernels gemma_rms_norm / fused_add_gemma_rms_norm ops, passing the raw (bf16/fp16) weight so the (1 + weight) offset and fp32 multiply are folded into the kernel (matching forward_native numerics).
• Falls back to forward_native when the kernels are unavailable in the installed package.
• Depends on [Qwen3.5] feat(layernorm): add fused XPU GemmaRMSNorm kernels vllm-xpu-kernels#532
• Also add the corresponding _custom_ops wrappers and guarded register_fake entries.

AI assistance (GitHub Copilot CLI) was used to author this change.

Test Result

Performance (enforce-eager prefill on Intel B70 input_len=8192)

Model	TP	eager (native fallback)	eager (fused)	torch.compile
Qwen3.5-9B	1	1.256 s	1.023 s	1.013 s
Qwen3.5-35B-A3B	4	0.580 s	0.419 s	0.421 s

Accuracy (GSM8K, lm-eval-harness, max_gen_toks=4096)

Model	TP	exact_match (strict)	exact_match (flexible)
Qwen3.5-9B	1	0.8923	0.8901
Qwen3.6-27B	4	0.6748	0.6755
Qwen3.5-35B-A3B	4	0.9462	0.9538

[claude]

Claude Code Review

This pull request is from a fork — automated review is disabled. A repository maintainer can comment @claude review to run a one-time review.

[jikunshang]

/ci run

[github-actions]

✅ Triggered Buildkite CI #86370 for commit 5dda576fa86c.

[jikunshang]

there are some failed case related to gemma in intel/ci, please check.
Multi-Modal Models (Standard) 2: qwen3 + gemma https://buildkite.com/vllm/intel-ci/builds/10131#01a05851-ff34-4e76-9a97-2d1b2f1d1fa3

[ccrhx4]

there are some failed case related to gemma in intel/ci, please check. Multi-Modal Models (Standard) 2: qwen3 + gemma https://buildkite.com/vllm/intel-ci/builds/10131#01a05851-ff34-4e76-9a97-2d1b2f1d1fa3

The CI failed is caused by the output tensor creation. Update the related code.

    # empty_like preserves x's strides, but the kernel requires a
    # contiguous out (unlike x, which it can handle non-contiguous).
    out = torch.empty(x.shape, device=x.device, dtype=x.dtype)
    ops.gemma_rms_norm(out, x, self.weight.data, self.variance_epsilon)

Now the failed CI seems unrelated to this PR. @jikunshang Would you please again kindly help? Thank you.

[jikunshang]

/ci run

[github-actions]

✅ Triggered Buildkite CI #86802 for commit 21cff155b1e5.