Add continuous batching support for concurrent LLM inference

[ronaldmannak]

Proposed changes

• Continuous batching engine that transparently serves multiple concurrent requests in a shared decode loop, with zero overhead for single requests
• BatchKVCache & BatchRotatingKVCache implementations with left-padding strategy for batched attention across sequences of different lengths
• InferenceScheduler actor that starts in single-request mode (TokenIterator) and automatically upgrades to BatchTokenIterator when a second request arrives; 3rd+ requests join the existing batch on the fly
• LRUPromptCache with trie-based prefix matching and LRU eviction for reusing KV state across requests with shared prefixes
• Batch-aware RoPE via the new applyRotaryPosition helper — all 45 modified MLXLLM model files with RoPE-based attention paths now use batch-aware position handling that works in both single and batched modes
• Automatic fallback for incompatible requests (VLMs, hybrid/SSM models using MambaCache or CacheList, and quantized KV-cache requests) — routed to the single-request path with no caller changes needed

Design

This PR adds batching support for text-generation LLMs. VLM requests are not batched in this PR, and embedding models are out of scope for this batching path.

Most LLM model types are supported. Batched generation is currently limited to models whose cache stack uses KVCacheSimple and/or RotatingKVCache, and is unavailable for models that use MambaCache or CacheList. Requests using those models automatically fall back to the single inference path. Support for MambaCache and CacheList will be added in a separate PR.

QuantizedKVCache is also not supported for batching. Requests with kvBits != nil, or requests that already carry quantized KV cache state, automatically fall back to the single inference path.

The batching system is opt-in via ModelContainer.scheduler:

let container = ModelContainer(context: context)
container.scheduler = InferenceScheduler()
// Existing generate() calls now batch transparently when the request is compatible

Incompatible models that use MambaCache and/or CacheList are:

• falcon_h1
• baichuan_m1
• lfm2
• lfm2_moe
• granitemoehybrid
• qwen3_next
• qwen3_5, qwen3_5_text, qwen3_5_moe
• nemotron_h
• jamba_3b

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code / installed pre-commit prior to committing changes
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)