Add first-class Qwen3.8 DFlash2 backend

[Deviad]

Summary

• add a first-class MLX dflash2 backend for Qwen3.8-27B bundles
• keep target and DFlash2 weights namespaced under one fail-closed artifact contract
• route ask, quickstart, and serve through the official dflash==0.1.0 implementation, with explicit target-only AR support
• support unquantized, 8-bit, and 4-bit draft loading, telemetry, local sidecars, Homebrew environments, and OpenAI-compatible serving
• add focused artifact/runtime/server/exactness tests and user documentation; no llama.cpp dependency

Verification

• uv lock --check
• Ruff: new DFlash2 source and tests pass; added-line scan reports no new Ruff findings in modified legacy files
• focused DFlash2 + related regression suite passes; real-model gates skip unless explicitly enabled
• unchanged artifact suite: 97 tests pass
• full server suite: 329 tests pass
• Python wheel and sdist build successfully; the wheel contains both new runtime modules and the dflash2 extra
• Homebrew-routed ask completed through DFlash2; quickstart --dry-run --json selected backend_id=dflash2; a live OpenAI-compatible server request returned HTTP 200
• 1,024-token prompt / 64-token greedy exactness passed for unquantized, 8-bit, and 4-bit DFlash2 plus native MTP, all with the same committed-token SHA-256
• selected 4-bit draft: 19.29747831754572 wall-clock tok/s versus 11.05228273050392 target-only AR and 9.804179514121502 native MTP on the same prompt and budget

Five existing tests/test_public_cli.py failures were reproduced unchanged on the base revision and are not regressions from this change.

Closes #308