Add mid-batch extend for text-only MLLM requests

[janhilgard]

Summary

• MLLM batch generator previously only started new batches when num_active == 0, serializing all concurrent requests. A short text-only request had to wait for a long request to finish generating all its tokens.
• Now text-only requests (no images/videos) can join an active batch mid-generation via MLLMBatch.extend(). Multimodal requests still wait for batch completion due to vision encoding shape constraints.
• This leverages the existing MLLMBatch.extend() method and BatchKVCache.extend() — no new infrastructure needed.

Before

Request 1 (long):  |===prefill===|=========generation (2000 tokens)=========|
Request 2 (short): |                    waiting...                           |===prefill+gen===|

After

Request 1 (long):  |===prefill===|=========generation (2000 tokens)=========|
Request 2 (short):        |==prefill+gen==| done!

Tested on Qwen3.5-35B-A3B (--mllm --continuous-batching)

• Long request: 53.8s (2000 max_tokens)
• Short request: 1.0s (joined mid-batch, finished 48.8s earlier)
• Both outputs correct, no artifacts

Test plan

• Concurrent text-only requests: short request joins active batch and finishes independently
• Output correctness: both requests produce coherent text
• Multimodal requests still wait for num_active == 0 (no regression)
• Edge case: _process_prompts returns None (handled, no extend called)

🤖 Generated with Claude Code

[janhilgard]

@Thump604 Superseded — mid-batch extend for text-only MLLM requests is already in main via #278. Closing.