Support mixed-precision and hybrid model configurations

[martingiovannigeyer]

Motivation

Support mixed-precision compressed-tensors checkpoints and the hybrid models covered by this patch without checkpoint-specific runtime forks.

Changes

• parse and preserve each compressed-tensors config groups own format
• retain exact ignore scoping and targeted ParallelLMHead quantization
• fix Ling lightning-attention and Nemotron Puzzle initialization
• fix compressed NVFP4 MoE W13 loading for FlashInfer CUTLASS by forwarding the per-layer [up, gate] ordering contract through the compressed-tensors wrapper

The last change is intentionally narrow. The NVIDIA ModelOpt method already exposes the same ordering contract; the compressed-tensors wrapper previously hid it and loaded [gate, up] into a CUTLASS kernel expecting [up, gate].

Validation

• 9 focused CPU regression tests pass in the pinned SGLang CUDA 13 runtime
• Qwen3.6 27B Unsloth NVFP4 repeated 1,000/1,000 outputs on RTX PRO 6000 and RTX 5090
• before the W13 fix, Qwen3.6 35B Unsloth NVFP4 agreement was roughly 42.5% with both the first patch commit and the three-commit patch head
• after 2b0dbfd9b8, 35B agreement is 95.7-95.8% on RTX PRO 6000 and 96.2% on RTX 5090
• full 1,000-request repeats completed on both GPUs with no MTP/speculative decoding
• Ling and Nemotron initialization regressions remain covered

No new checkpoint format, kernel, runtime flag, or forward hot path is introduced.

[martingiovannigeyer]

Superseded by two scoped PRs: one for mixed-precision/NVFP4 support and one for the independent Ling/Nemotron initialization fix.