perf: token-level multimodal mm token expansion.

[kongweiqian]

Description

Add token-level multimodal placeholder expansion for VLM prompt processors (Qwen2-VL, Qwen3-VL, GLM4V).

The legacy path performs string-level placeholder insertion followed by a full re-tokenization of the entire prompt. The new path operates directly on the token sequence: it locates placeholder token IDs and expands them in-place with the correct image/video token count, avoiding redundant tokenization work.

This reduces tokenization_latency_seconds for multimodal requests—especially those with high-resolution or multi-image inputs—while producing identical token_ids and span metadata.

A new mm_prompt_expansion_latency_seconds counter is added to track the expansion step independently.

Related Issues

N/A

Change Type

• Bug fix
• New feature
• Performance improvement
• Refactor
• Documentation
• Test
• Build or CI

Pull Request Checklist

PR Title and Commit Messages

• The PR title and each commit message follow the xLLM commit format: <type>: <subject>.

Pre-commit Checks

• I have installed pre-commit by running pip install pre-commit or an equivalent command.
• I have installed the hooks with pre-commit install.
• I have run pre-commit run --all-files and fixed any reported issues.

Self Review

• I have self-reviewed the code according to .agents/skills/code-review/references/custom-code-style.md, especially code written or assisted by AI.
• I have rebased this PR onto the latest main branch.

Build and Test Coverage

• Tests have been added or updated as needed.
• CUDA: python setup.py build test has passed on a CUDA machine.
• NPU: python setup.py build test has passed on an NPU machine.
• MLU: python setup.py build test has passed on an MLU machine.

Reviewer Notes

This PR touches xllm/processors/ (prompt processors for GLM4V, Qwen2-VL, Qwen3-VL) and xllm/core/common/metrics.*.

Validated on NPU (Ascend 910) with Qwen2.5-VL-3B:

• Correctness: new token-level path produces identical token_ids and multimodal span offsets vs. legacy string-expand path across 256×256 to 2048×2048 resolutions.
• Performance: up to 4× reduction in tokenization latency for 2048×2048 images; no regression for text-only prompts.
• Multi-image ordering preserved (verified with 5-image interleaved prompts).