[Bugfix][SM70] Repair #345 Qwen4Exp integration gates

[yangzhuxinyzx]

Purpose

Repair the latest-main source and static-gate issues found while auditing #345 without changing its model policy or performance defaults.

• Preserve stacked QKV/merged-column loader metadata while failing clearly on unknown parameters.
• Restore the inherited Qwen attention output contract so both QSA and ordinary full-attention paths are callable.
• Correct heterogeneous Mamba state shape/dtype annotations.
• Narrow the speculative config type and align Qwen4Exp PLE device management with torch.accelerator.
• Fix changed-test type annotations and forbidden torch.cuda synchronization calls.

Test Plan

• Run every changed-file pre-commit hook.
• Compile all [Model][SM70] Adapt Qwen3.8 Flash Next NVFP4 #345 Python files.
• Run the focused Qwen4Exp config/model/PLE/QSA/weight-loading, stacked-loader, custom-AR, KV-cache, spec-decode, and V2 worker CPU contracts.

Test Result

• All changed-file pre-commit hooks passed, including ruff, mypy, forbidden-import and accelerator API gates.
• Python compile passed for all [Model][SM70] Adapt Qwen3.8 Flash Next NVFP4 #345 Python files.
• Focused source/CPU contracts: 74 passed, 3 CUDA-only skipped.
• The two locally attempted CUDA-only cases were not used as acceptance evidence: all eight V100s are occupied by an unrelated PP2/TP4 service and this audit worktree has no matching _C binary. No process was stopped or preempted.
• Existing [Model][SM70] Adapt Qwen3.8 Flash Next NVFP4 #345 recorded V100/model evidence remains unchanged.

Audited #345 head: ceb543c05573c9bcbf1cc9563bab0c2e8c746ae3.
Repair head: 95fdf660b325ef1627a01b414518d1c45f7e67d0.