Harden legacy converter checkpoint loading

[igorsantosgrueiro]

Summary

• Load legacy OpenNMT-py and Fairseq converter checkpoints with weights_only=True by default.
• Add --unsafe_deserialization as an explicit opt-in for trusted legacy checkpoints that still require Python pickle globals.
• Keep this hardening scoped to the two deprecated converter paths that explicitly opted out of PyTorch's weights-only loader.

This preserves a compatibility path for existing trusted local checkpoints while avoiding unsafe pickle deserialization by default.

Fixes #2034.

Rationale

Some legacy OpenNMT-py/Fairseq checkpoints may include Python objects in addition to tensors and may not load with weights_only=True. Instead of removing that behavior entirely, this PR makes the unsafe behavior explicit and documented at the converter CLI boundary.

Testing

• Not run locally: upstream checkout and optional converter dependencies are not available in this environment.