feat: integrate RDMA support with MLX backend

[localai-bot]

This PR adds RDMA support to the MLX backend in LocalAI, enabling distributed inference across Apple Silicon machines.

Summary of Changes

Backend Enhancements (backend/python/mlx/backend.py)

• Added _initialize_rdma() method to initialize mlx.distributed when MLX_GRPC_SERVERS is set
• RDMA workers are identified via environment variable (no hostfile needed, manual worker coordination via P2P)
• Model loading and generation logic now supports collective operations when RDMA is active

CLI Worker Support (core/cli/worker/worker_mlx.go)

• New mlx-rdma worker type for launching MLX backends in distributed mode
• Follows the same pattern as llama.cpp workers (no P2P registration, uses MLX_GRPC_SERVERS)

P2P Integration (core/cli/run.go)

• Sets MLX_GRPC_SERVERS environment variable alongside LLAMACPP_GRPC_SERVERS in TunnelCallback
• Enables automatic worker discovery and IP collection for MLX RDMA workers

Design Decisions

1. No hostfile needed: Since workers are launched manually via local-ai worker mlx-rdma, we don't need automatic process spawning via mlx.launch --hostfile
2. Environment variable-based: Uses MLX_GRPC_SERVERS (same as LLAMACPP_GRPC_SERVERS) to pass worker IPs to the backend
3. P2P reuse: Leverages existing LocalAI P2P infrastructure for worker discovery, but backend handles RDMA coordination

Testing

1. Launch main instance with P2P enabled: local-ai run --p2p --token <token>
2. Launch workers on each node: local-ai worker mlx-rdma
3. Workers register via P2P; main instance sets MLX_GRPC_SERVERS env var
4. MLX backend initializes RDMA when MLX_GRPC_SERVERS is set

Notes

• Requires mlx with JACCL backend support (mlx-jaccl-cluster integration)
• Current implementation assumes all workers have identical model paths
• Future work: Add model sharding via model.shard(mx.distributed.world_size())