Add native DSpark speculative decoding

[Blaizzy]

Summary

• add native DSpark checkpoint loading and speculative decoding for RadixArk/Qwen3.8-27B-DSpark
• auto-detect DSpark checkpoints and expose --draft-kind dspark plus load-time --draft-bits quantization
• add acceptance-adaptive draft depth while keeping explicit block sizes fixed for reproducible tuning
• make speculative prefill sizing depend on live Metal headroom instead of a specific Mac tier
• document CLI and server usage for Qwen3.8

Root cause and memory fix

Qwen3.5/3.8 previously coupled hidden-layer capture with Gated DeltaNet rollback capture. DSpark prefill only needs five hidden-state taps, but a 1k prompt retained per-token rollback intermediates across all 48 linear-attention layers. This could require more than 150 GB before other activations and caused unchunked prefill to be killed.

This change separates hidden capture from rollback-state capture. Rollback intermediates are now retained only for short speculative verification rounds. It also explicitly collects and clears the dense BF16 drafter tensors after load-time quantization.

Impact

On an Apple M5 Max with 48 GB unified memory, using Qwen3.8-27B-4bit with a fixed three-proposal DSpark block:

• 1,029-token unchunked prefill: 857.6 tok/s
• 64-token decode: 43.57 tok/s
• peak Metal memory: 19.54 GB
• baseline decode: ~33.6 tok/s (about 30% slower)
• the same unchunked prefill previously OOMed

The quantized drafter's live parameter footprint is about 0.765 GB after reclaiming the dense load-time copy.

Validation

• 162 passed, 1 deselected in mlx_vlm/tests/test_speculative.py
• 31 passed targeted server and prompt tests
• targeted Ruff fatal/error checks passed
• git diff --check passed
• real-model 1k-input/64-output server benchmark passed

The deselected speculative test is the existing exact-BF16 quantized-linear equality check; it remains unrelated to DSpark and differs only in numerical accumulation.