Add local MLX backend support and benchmark tooling

[i386]

What this adds for users

This PR adds local mlx backend support alongside the existing llama.cpp path.

Backend behavior

• Local MLX model directories now run through the mlx backend automatically.
• GGUF files continue to run through the llama backend.
• mesh-llm can launch mlx_lm.server either from PATH, via python -m mlx_lm.server, or through --mlx-server-bin / MESH_LLM_MLX_SERVER_BIN.
• Hugging Face snapshot paths now normalize to stable repo-style model ids instead of snapshot hashes.
• Local MLX requests rewrite the upstream model field to the actual model directory so the mesh-facing alias stays stable while the runtime still accepts the request.
• Backend election and startup now treat MLX as solo-only for now, skipping RPC split mode when the backend does not support it.

Docs and operator workflow

• The README now explains that local MLX model folders are supported directly.
• evals/backend-benchmark.py was expanded into a generic local backend benchmark so the same harness can compare matched exports across different backends.
• evals/README.md now documents both the simple llama/MLX workflow and the JSON-spec workflow for arbitrary backend combinations.

Benchmarking

• The benchmark now accepts paired llama and MLX exports for the same model family.
• Model identity normalization was tightened so quantized suffixes such as Q4_K_M and 4bit collapse to the same comparison key.
• The current benchmark results were collected locally on macOS and should be treated as machine-specific rather than universal.

Current benchmark results

Paired-export benchmark using Qwen2.5-0.5B-Instruct:

• mlx: local 4bit MLX export
• llama: local Q4_K_M GGUF export
• concurrency sweep: 1,4,8
• iterations: 3
• warmup: 1
• max tokens: 64

backend  llama  startup_s  6.706  model_id  Qwen2.5-0.5B-Instruct-Q4_K_M
concurrency  avg_ttft_s  avg_total_s  avg_batch_wall_s  avg_decode_tok_s  avg_overall_tok_s  avg_batch_overall_tok_s
          1       0.016        0.684             0.687            95.915             93.630                   93.161
          4       0.045        1.012             1.017            66.228             63.308                  251.875
          8       0.512        1.470             1.963            66.842             48.946                  260.839

backend  mlx  startup_s  6.031  model_id  Qwen2.5-0.5B-Instruct-4bit
concurrency  avg_ttft_s  avg_total_s  avg_batch_wall_s  avg_decode_tok_s  avg_overall_tok_s  avg_batch_overall_tok_s
          1       0.103        0.520             0.520           153.386            123.120                  123.067
          4       0.394        1.471             1.477            59.463             43.544                  173.495
          8       0.881        2.380             2.488            46.009             29.898                  206.845

Key read from that run:

• MLX wins single-request throughput and total time on this machine.
• llama keeps a much lower TTFT at low concurrency.
• llama scales better once concurrency increases.

Validation

• just build
• cargo test -p mesh-llm
• cargo check -p mesh-llm
• python3 -m py_compile evals/backend-benchmark.py
• Local paired-export benchmark using Qwen2.5-0.5B-Instruct in f16/bf16 and Q4_K_M/4bit forms

[michaelneale]

super excited for this one!

[michaelneale]

I think can retire this one in favour of a native one: #103