Add MLX backend for raw / L0 / L1

[TerenceLiu98]

## Summary

This PR adds an optional MLX-LM backend for running AnyJev on Apple Silicon.

The backend supports raw, L0 and L1 decisions through direct next-token logits. L2 is intentionally not supported because the backend does not expose a stable hidden-state interface.

## Changes

- Add `MLXBackend` under `anyjev/backends/mlx.py`
- Add the optional `mlx` dependency extra
- Add `scripts/mlx_parity.py` for comparison with `HFBackend`
- Add unit tests covering:
  - batched inference
  - right-padding
  - last valid token position
  - restricted log-probability extraction
  - invalid prompt inputs
- Add an `@pytest.mark.engine` smoke test for a real MLX-LM model
- Document installation and usage in `README.md`
- Record MLX and MLX-LM licensing in `THIRD_PARTY.md`
- Add a changelog entry

## Usage

```bash
pip install -e ".[mlx]"

from anyjev import Decider
from anyjev.backends.mlx import MLXBackend

backend = MLXBackend(
    "Qwen/Qwen3-0.6B-MLX-4bit"
)
decider = Decider(backend, level="L0")

Validation

• ruff check anyjev bench demo scripts space tests
• pytest -q -m 'not engine' --ignore=tests/test_truncate_and_vllm.py
• pytest -q -m engine tests/test_mlx_backend.py

The MLX engine test is skipped when MLX, MLX-LM, Metal, or the configured model is unavailable.

No benchmark numbers are added in this PR. The parity script reports differences when run with matching Transformers and MLX models.

[MorrisZJ]

Thanks @TerenceLiu98, and sorry for the wait; 0.3.0 landed in between. We ran this on Apple Silicon (M5, mlx 0.32.3, mlx-lm 0.32.0, Qwen2.5-0.5B-Instruct): your unit tests and the engine smoke test pass; scripts/mlx_parity.py against transformers on CPU (fp32) gives 15/15 argmax agreement (max |Δ| 0.17 restricted log-softmax, MLX in bf16); and L0 decisions through Decider match on 23/24 choice, yes/no and score items (max |Δp| 0.03; the one flip is a near-tie). So the readout is right. A few changes before we merge:

1. Logits at the last position only. self.model(batch) materialises full-vocabulary logits at every position before the last one is selected. Measured at batch 8 × ~1,800 tokens: peak 6.29 GB, against 1.95 GB when the backbone runs first and the LM head is applied to the last position only (h = model.model(x)[rows, lengths - 1], then model.lm_head(h), or model.model.embed_tokens.as_linear(h) for tied embeddings). The output is bit-identical and about 30% faster, and the gap grows with the model and the prompt, which matters on 8–16 GB Macs. Please fall back to the full forward if a model exposes neither.
2. The parity script can't run on a Mac. HFBackend defaults to device="cuda"; please add --hf-device (default cpu, with float32).
3. README: 0.3.0 rewrote it and the section this edits no longer exists. Please drop the README change; the docstring is enough.
4. 0.3.0 removed L1: "raw / L0" in the title, docstring and CHANGELOG.
5. Rebase onto main (CHANGELOG.md, pyproject.toml, THIRD_PARTY.md), and note the mlx / mlx-lm versions you checked in the docstring.

We'll rerun the same checks and merge.