Trust resolved draft config for Gemma4 MTP assistant loading (#597)

This PR is:

- To use vLLM’s resolved `draft_model_config` as the source of truth for
Gemma4 MTP assistant loading.
- To remove fallback reads from the outer speculative config.
- To keep the assistant cache key tied to the draft model revision vLLM
actually resolved.

vLLM 0.27 builds the draft `ModelConfig` before Metal sees the
speculative config, so the Gemma4 assistant loader should trust that
typed contract directly.

Local smoke prompts confirmed the Gemma4 target and MTP assistant load
together and produce coherent chat outputs on Metal.

| Prompt | Output |
|---|---|
| `Write one short sentence about the Moon.` | `The Moon orbits the
Earth, casting a silvery glow across the night sky.` |
| `Answer with one word: What color is grass?` | `Green` |
| `Complete this sentence naturally: Apple pie tastes` | `Here are
several ways to complete the sentence, depending on the desired tone and
focus: ...` |

Signed-off-by: Yuan Lik Xun <lxyuan0420@gmail.com>