fix(gemma4): load mixed compressed-tensors nvfp4

[Poyo-Yon]

Fixes #492

Summary

Teach the Gemma 4 checkpoint reader to load the mixed compressed-tensors layout used by unsloth/gemma-4-26B-A4B-it-NVFP4:

• fuse per-channel FP8 q/k/v weight_scale roles alongside their weights and normalize [out_features, 1] bf16 scales to FreeToken's flat fp32 buffers;
• map compressed-tensors NVFP4 shared-MLP roles through the existing nvfp4_parts_ct / ct_nvfp4_fuse helpers;
• build the NVFP4 routed-expert source spec from the installed checkpoint dialect, including reciprocal compressed-tensors global scales;
• ignore static KV quantizer metadata which is not part of the runtime KV cache.

The existing ModelOpt Gemma 4 path is retained unchanged.

Tests

Weight-free regression coverage:

pytest -q \
  tests/models/test_gemma4_compressed_tensors.py \
  tests/models/test_gemma4_vision.py \
  tests/models/test_gemma4_gguf_rope.py

6 passed, 4 skipped in 20.93s

Real-hardware validation:

• OS: Windows 11 + Ubuntu 24.04 under WSL2
• GPU: NVIDIA GeForce RTX 5060 Ti 16 GB, driver 616.92
• Other GPU: NVIDIA GeForce RTX 5070 Ti 16 GB
• FreeToken base: cac247a (0.1.3)
• Checkpoint: unsloth/gemma-4-26B-A4B-it-NVFP4
• Mode: text-only, 4096 context, NVFP4 expert offload, 40% expert cache
• Result: model load, serial expert-bank build, 52,597-token KV-cache allocation, CUDA graph capture, /v1/models, and /v1/chat/completions all completed
• Three post-warm-up requests: 34.48, 59.44, and 60.65 output tok/s (median 59.44 tok/s)

Exact serve command is included in the linked issue.

AI assistance was used to prepare the patch and tests. The contributor reviewed the behavior and ran the real-checkpoint validation described above.