fix: surface deep_gemm load failures instead of silently disabling FP8

[S1ro1]

Problem

fp8_linear.py wraps the deep_gemm import in a bare except ImportError:

try:
    import deep_gemm
except ImportError:
    deep_gemm = None

This is meant to support CPU-only environments where deep_gemm isn't installed. But ImportError also fires when deep_gemm is installed and its native extension fails to load — e.g. a wheel that links libcudart.so.13 running on a host with only CUDA 12. In that case the symbol is silently set to None, and the real failure doesn't surface until the first FP8 forward pass, as a misleading:

AttributeError: 'NoneType' object has no attribute 'fp8_gemm_nt'

We hit this on a multi-node GLM-5.1 run: the deep_gemm release wheel linked CUDA 13 libs while the nodes shipped CUDA 12.9, so every FP8 matmul died with the NoneType error deep in the trainer — hours of misdirection before tracing it back to a swallowed libcudart.so.13: cannot open shared object file at import.

Fix

Narrow the except to ModuleNotFoundError:

• Not installed (CPU-only) → still raises ModuleNotFoundError, falls back to None. Unchanged behavior.
• Installed but failed to load (missing/mismatched CUDA libs) → raises a plain ImportError, which now propagates with its real cause, making the misconfiguration obvious at import time.

Aligns with the "errors should never pass silently" principle — the CPU fallback stays, but a broken native install no longer masquerades as a missing-attribute bug.