Fix MRoPE dimension expansion in vLLM adapter for hybrid Qwen models

[xuefgu]

Description

When serving models that use MRoPE — such as Qwen3.5 and Qwen3-Omni (use_mrope=True) — vLLM passes 3D position indices of shape (3, num_tokens).

The vLLM adapter was expanding this at axis=1, producing a shape of (3, 1, num_tokens). When broadcasted against key projections during RoPE application, it incorrectly inflated the key sequence dimension (resulting in (batch, max_num_tokens, heads, head_dim)), ultimately leading to a shape validation failure in the RPA kernel: ValueError: Expected k.shape=(33554432, 1, 256) to be equal to v.shape=(2048, 1, 256).

The solution expand at axis=-1 to obtain the correct shape for both MRoPE and non-MRoPE cases.

Tests

Manual tests in E2E runs.

Checklist

Before submitting this PR, please make sure (put X in square brackets):

• I have performed a self-review of my code. For an optional AI review, add the gemini-review label.
• I have necessary comments in my code, particularly in hard-to-understand areas.
• I have run end-to-end tests tests and provided workload links above if applicable.
• I have made or will make corresponding changes to the doc if needed, including adding new documentation pages to the relevant Table of Contents (toctree directive) as explained in our documentation.

[codecov]

Codecov Report

❌ Patch coverage is 0% with 1 line in your changes missing coverage. Please review.

Files with missing lines	Patch %	Lines
...t/integration/vllm/maxtext_vllm_adapter/adapter.py	0.00%	1 Missing ⚠️

📢 Thoughts on this report? Let us know!