Fix candidate wheel assembly

[yyj6666667]

What does this PR do?

Fix candidate wheel assembly in the existing release workflow. No training runtime or kernel changes.

• Accept CPU-only KT binaries while retaining CUDA architecture checks.
• Recognize the wheel's own metadata rather than vendored metadata.
• Preserve wheel-relative native dependencies in lazy payloads.
• Use the existing Accelerate carrier capacity and exclude CUDA libraries supplied by PyTorch.

The payload layout requires the companion loader change in kvcache-ai/sglang#96. This replaces the packaging portion of #2203; unrelated changes are excluded.

Tests

python3 -m pytest -q .github/release: 61 passed.

• Fresh six-variant CPU and CUDA builds passed native dependency / architecture audits; all five assembled wheels pass size, RECORD, and twine checks.
• Two fresh Python 3.12 environments passed normal hash-locked installation, pip check, and installed-byte verification.
• Qwen3-30B-A3B and native FP8 DeepSeek-V3.1 LoRA: two-GPU training, ordinary/expert updates and full checkpoints passed.
• Kimi: eight-GPU training and fresh-process resume passed exact adapter, optimizer, scheduler, RNG and step-loss comparisons; wheel-only SGLang LoRA loading and generation passed.
• Qwen and public FP8 GLM-5.3-Flash generation and JSON-schema checks passed. Full Neko style validation remains pending.

This PR does not publish wheels.