Add Metal memory stats to MLLMScheduler

[janhilgard]

Summary

• MLLMScheduler.get_stats() was missing Metal memory metrics (metal_active_memory_gb, metal_peak_memory_gb, metal_cache_memory_gb) that the regular Scheduler already reports (scheduler.py:1818-1822)
• SimpleEngine.get_stats() was also missing Metal memory metrics entirely
• BatchedEngine.get_stats() nested MLLM scheduler stats under a mllm_scheduler key, but /v1/status looks for metal_* at top-level — so Metal stats were always null for vision models even after the scheduler fix
• All three issues together caused monitoring dashboards to show no Metal memory data for vision models (e.g. Qwen3-VL on BatchedEngine)

Changes

1. mllm_scheduler.py — Added import mlx.core as mx and Metal memory stats block to get_stats() (identical to scheduler.py:1818-1822)
2. engine/simple.py — Added Metal memory stats to SimpleEngine.get_stats()
3. engine/batched.py — Promote metal_*_memory_gb keys from nested mllm_scheduler dict to top-level, so /v1/status can find them (consistent with how non-MLLM stats are merged via update())

Test plan

• Verify curl http://localhost:1236/v1/status returns non-null metal.active_memory_gb, metal.peak_memory_gb, metal.cache_memory_gb for an MLLM model (Qwen3-VL-8B — confirmed 5.76 GB)
• Verify no regression on regular LLM scheduler stats

🤖 Generated with Claude Code