feat: Enable continuous batching for MLLM models

[lubauss]

Summary

• Add MLLMModelWrapper to extract logits from LanguageModelOutput objects, making MLLM models compatible with BatchGenerator
• Fix tokenizer handling in scheduler to work with processors (e.g., Qwen3VLProcessor) that wrap the actual tokenizer
• Fix _get_stop_tokens() to check both processor and nested tokenizer for EOS tokens

Problem

Continuous batching (with prefix caching) was broken for multimodal LLM models like Qwen3-VL and Gemma 3:

1. AttributeError: 'Qwen3VLProcessor' object has no attribute 'encode' - The scheduler called tokenizer.encode() but MLLM processors don't have an encode method directly
2. 'LanguageModelOutput' object is not subscriptable - BatchGenerator expected raw logits array but MLLM models return LanguageModelOutput(logits=...) objects

Solution

1. Created MLLMModelWrapper that wraps MLLM models and extracts .logits from output
2. Added _get_actual_tokenizer() to extract the nested tokenizer from processors
3. Added _decode_tokens() helper to use the actual tokenizer for decoding
4. Updated tokenize logic to check for tokenizer.tokenizer.encode() for processors

Performance Results

Tested on M4 Max 128GB with mlx-community/Qwen3-VL-30B-A3B-Instruct-4bit:

Request	Latency	Speedup
1 (cold)	21.7s	-
2 (cached)	1.15s	19x
3 (cached)	0.79s	28x
4 (cached)	0.78s	28x

Multi-turn conversation (6 turns with 17K token context):

• Average latency improvement: 90.7%
• Overall speedup: 10.76x

Test plan

• Tested with Qwen3-VL-30B-A3B-Instruct-4bit model
• Verified prefix caching works across multi-turn conversations
• Confirmed standard LLM models still work (no regression)
• Tested both streaming and non-streaming modes

🤖 Generated with Claude Code