fix(models): support nvidia/Qwen3.8-Flash-Next-NVFP4 in qwen4_exp

[Eng-Ahmd]

Problem: nvidia/Qwen3.8-Flash-Next-NVFP4 fails to load. First with missing MoE expert sources on every layer, and after working around the quant flag, with KeyError on model.layers.0.linear_attn.out_proj.weight_scale. RadixArk/Qwen3.8-Flash-Next-NVFP4 loads fine.

Root cause (python/freetoken/models/qwen4_exp/config.py):

• Expert detection only checks the top-level quant_algo for fp4. The nvidia checkpoint labels it MIXED_PRECISION and carries the NVFP4 experts in quantized_layers/config_groups, so experts resolved to none and fell into the BF16 offload path.
• attn_quant probes layers.0.self_attn.q_proj, but layer 0 is linear_attention and has no self_attn module. The ignore match always misses, so attn resolved to nvfp4 and GDN out_proj was built as NVFP4 while the checkpoint stores it BF16.

Change:

• Add _is_nvfp4_experts(): top-level fp4, or mixed plus an FP4 experts entry in quantized_layers or a 4-bit float experts group in config_groups.
• Compute layer split before quant parsing. Probe self_attn on full_ids[0] and linear_attn on linear_ids[0]; attn stays none unless either side is actually quantized. Empty groups are treated as ignored.
• docs/models.md: list the nvidia checkpoint. weight.py docstring updated (loader regex already matched both layouts).

Verification:

• tests/models/qwen4_exp/test_config.py: 8 passed.
• tests/models/qwen4_exp/test_weight_ckpt.py against a local nvidia/Qwen3.8-Flash-Next-NVFP4 copy: 41 passed (dense names equal the model state dict, sampled tensors match).
• parse_config on the nvidia checkpoint now yields expert=nvfp4 attn=none dense=none lm_head=none; RadixArk synthetic config unchanged.

Scope: qwen4_exp only. No Laguna changes.