fix: make Nemotron-H import error adaption on Megatron-Core 0.16.1

[hazelduan]

Make Nemotron-H registration conditional when the installed Megatron-Core does not provide megatron.core.models.hybrid. On Ascend A3 with Megatron-Core 0.16.1 and USE_MCORE_GDN=0, the training initialization import passed and Nemotron-H was correctly skipped.

[hjh0119]

Thanks for the fix!

Instead of guarding the import at the init.py level, would it be cleaner to wrap the relevant imports inside nemotron_h.py itself with a try/except ImportError?

This way the file is self-contained — it won't break on import regardless of where it's imported from, and we don't need to maintain a whitelist in init.py every time a new model has optional dependencies.

[hazelduan]

Thanks for the suggestion — addressed in 04e21c5. The HybridModel availability check now lives in nemotron_h.py, while gpts/__init__.py keeps its regular import list. I kept the exception guard narrow so only failures originating from megatron.core.models.hybrid skip registration; unrelated import errors are still raised. Verified on A3 with Megatron-Core 0.16.1: package/direct module imports pass, Nemotron-H remains unregistered when HybridModel is absent, and lint passes.