feat(mlx): support Qwen3.8 Flash Next (qwen4_exp)

[tharalds]

Motivation

Qwen3.8 Flash Next uses the qwen4_exp architecture, whose QSA/Gated
DeltaNet cache, hyper-connections, n-gram PLE storage, and weight sanitizer
currently live in MLX-VLM rather than MLX-LM. EXO could discover the model but
could not load or generate from a compatible checkpoint through its MLX engine.

Changes

• Pin MLX-VLM to an immutable upstream revision containing Qwen4-Exp QSA/APC,
  continuous-batching, MTP, and external-PLE support.
• Route qwen4_exp through MLX-VLM while preserving MLX-LM loading for every
  existing architecture.
• Adapt MLX-VLM's structured language-model output to EXO's logits contract.
• Add MLX-VLM cache types to EXO's cache model, including detached recurrent
  snapshots, QSA trimming, prefill rollback, and prefix-cache reuse.
• Repair Qwen4-Exp hybrid layer/cache indices after pipeline slicing. Tensor
  parallel placement remains disabled until QSA/GDN/MoE/PLE sharding has a
  dedicated bit-exact implementation.
• Normalize Qwen3.8 reasoning efforts to its supported low, medium, and
  xhigh template values and recognize its EOS tokens.
• Add a conservative catalog card for the corrected
  sh0wie/Qwen3.8-Flash-Next-REAP-288-MLX-4bit checkpoint. Vision is not
  advertised because the checkpoint publisher has only evaluated text after
  pruning.
• Add focused Metal-backed tests, model-card validation, a reusable full-weight
  smoke harness, and operator documentation.

The disaggregated-prefill wire path now rejects MLX-VLM caches explicitly
instead of attempting to serialize incomplete recurrent/QSA state.

Why It Works

MLX-VLM owns the current qwen4_exp implementation and sanitizes its official
conditional-generation weights. EXO retains only that model's language module,
wraps its structured output as logits, and continues to use EXO's existing chat,
generation, placement, and prefix-cache orchestration.

The recommended REAP-288 checkpoint corrects two source-conversion defects:
RMSNorm tensors use the zero-centered convention expected by MLX-VLM, and PLE
tensors plus per-tensor quantization overrides use the runtime's module names.
Its external PLE manifest also keeps active memory substantially below the
checkpoint's on-disk size.

Test Plan

Manual Testing

• Hardware: Mac Studio, Apple M3 Ultra, 512 GB unified memory.
• Downloaded all 131 indexed shards of the recommended checkpoint and verified
  that no indexed file was missing or empty.
• Native MLX-VLM returned the exact requested coherent response, peaking at
  41.7 GB active memory.
• EXO's loader, chat template, cache, and generator returned the exact requested
  coherent response, peaking at 44.5 GB active memory.
• Thinking mode produced the correct answer with a properly delimited reasoning
  section.
• A repeated full-weight EXO request reused 18 of 20 prompt tokens and dropped
  from 2.8 seconds to 0.7 seconds while returning identical correct output.
• An older conversion was tested through both native MLX-VLM and EXO and
  produced the same incoherent output; it is deliberately not cataloged.

Automated Testing

• 109 focused/regression tests passed (42 hardware/model-dependent tests
  deselected), covering Qwen4-Exp loading, real Metal prefill/decode, QSA and
  recurrent cache restore, pipeline handling, reasoning, API tool use, parsers,
  and disaggregated-prefill behavior.
• ruff check .
• ruff format --check .
• basedpyright — 0 errors, 0 warnings.
• uv lock --check