Fix Flash-Next FP8 n-gram scalar scales

[GrahamJenkins]

Summary

The stock converter fails on the official Qwen/Qwen3.8-Flash-Next-FP8 revision 236dfdf285828023ca3bcd3f37366c58a3469b13 because its 128 PLE n-gram FP8 shards share one BF16 scalar weight_scale rather than providing per-weight weight_scale_inv tensors.

This narrow fix falls back to that shared scalar only for n-gram shards, broadcasts it during FP8 dequantization, and accounts for the shared source tensor in fail-closed accounting. Ordinary block-scale weight_scale_inv conversion remains unchanged.

Validation

• Added synthetic regression coverage for both ordinary 128x128 block weight_scale_inv and official n-gram shared scalar weight_scale layouts.
• Targeted regression tests pass on Linux with a NumPy-backed MLX stand-in; native MLX execution is not available on this host.
• Completed the full 185.6 GB official-source conversion, Forge verification, and model-load verification on Nexus. The verified patched converter SHA-256 is 4ecc6eb37366739ba214778adea7ba6aeacf133817d5162cba684abc8f1a8b73.