Add native dense Llama generation to EMLXAxon

[hfiguera]

Adds an end to end native dense generation path for Llama 3.2 style checkpoints in emlx_axon.

This includes:

• dense Llama safetensors loading
• native Llama model state and generation loop
• an emlx_axon Llama compute plugin
• EMLX native ABI support needed by the plugin
• TextGeneration dispatch for Qwen3 and Llama states
• strict benchmark support for dense Llama validation
• focused loader, generation, native math, and TextGeneration tests

The goal is to keep the user facing model path in emlx_axon, while using native MLX support where it materially improves generation speed.

Benchmark was run on an Apple M4 Max with 128 GB RAM, using meta-llama/Llama-3.2-1B from local Hugging Face safetensors, f16, on the EMLX GPU backend.

The benchmark was run in strict length mode, so all paths generated exactly 64 tokens with the same prompt, tokenizer, generation config, device, precision, and greedy decoding.

Command:

cd emlx_axon
EMLX_AXON_LOCAL_EMLX=true \
EMLX_DENSE_LLAMA_MODEL=/path/to/local/Llama-3.2-1B/snapshot \
EMLX_DENSE_LLAMA_DEVICE=gpu \
EMLX_DENSE_LLAMA_TYPE=f16 \
EMLX_DENSE_LLAMA_SEQUENCE_LENGTH=128 \
EMLX_DENSE_LLAMA_MAX_NEW=64 \
EMLX_DENSE_LLAMA_RUNS=5 \
EMLX_DENSE_LLAMA_WARMUP_RUNS=1 \
EMLX_DENSE_LLAMA_STRICT_LENGTH=true \
EMLX_DENSE_LLAMA_JSON=/tmp/llama_dense_strict.json \
mix run bench/validate_llama_dense.exs

Results:

• stock Bumblebee + EMLX: 96.199 tok/s p50, 665.287 ms p50 duration
• EMLXAxon rewrite path: 130.486 tok/s p50, 490.473 ms p50 duration
• native dense Llama path: 172.203 tok/s p50, 371.654 ms p50 duration

All three paths generated 64 tokens in every measured run and all finished with length.

The native dense path is about 1.79x faster than stock Bumblebee + EMLX and about 1.32x faster than the existing EMLXAxon rewrite path in this strict comparison.

Validation:

cd emlx_axon
EMLX_AXON_LOCAL_EMLX=true EMLX_TEST_DEFAULT_GPU=true \
  mix test --exclude quantized_inference --warnings-as-errors

EMLX_AXON_LOCAL_EMLX=true \
  mix format --check-formatted mix.exs lib test bench

Result:

• 138 passed, 2 excluded
• formatting check passed

Note: CI uses the sibling emlx dependency through EMLX_AXON_LOCAL_EMLX=true because this PR adds the emlx native ABI support and the emlx_axon plugin that depends on it in the same branch. Before release, emlx_axon should depend on the next emlx version that contains this ABI.