Support loading pretrained weights from local checkpoints

[Jake Smith (jsmith13)]

Summary

• add --pretrained-model-path for local Fairseq checkpoints and raw model state dictionaries
• keep existing named checkpoint downloads and output-layer reset behavior unchanged
• reject ambiguous name-plus-path configuration and provide explicit missing or malformed checkpoint errors
• support the local path in the evaluation entry point and document trusted-source and distributed-filesystem requirements
• add focused loader tests

This PR is independent of #216 and can be reviewed or merged separately.

Validation

• 8 focused unit tests for local, named, invalid, and distributed-initialized loading paths
• exact PCQM4Mv2 GPU inference parity from a mounted local checkpoint
• output-layer reset verification with encoder weights retained
• two-process distributed load from a shared local checkpoint
• Fairseq CLI parsing for --pretrained-model-path