Fixes KeyError during Gemma3 LoRA fine-tuning initialization on Optimum Habana 1.18.1

[mkpatel3-github]

Root cause: PEFT removes LM head weight entry from Linear._parameters dict while leaving the attribute intact. When model moves to HPU and tie_weights() runs, PyTorch tries to re-register the parameter and hits the existing attribute.

Solution:

• Added safe fallback _safe_tie_weights() that manually re-ties embeddings without calling register_parameter
• Created _replace_module_parameter() helper that:
  • Overwrites existing _parameters[name] entries directly
  • Restores missing dict entries when attribute still exists as nn.Parameter
  • Deletes stale non-Parameter attributes before re-registering
  • Falls back to direct _parameters injection if register_parameter still fails
• Added diagnostic logging at each fallback step

Changes:

• optimum/habana/transformers/trainer.py:
  • Import OrderedDict
  • Wrap _move_model_to_device() tie_weights call in try/except
  • Add _safe_tie_weights() method
  • Add _replace_module_parameter() helper

Target: Optimum Habana 1.18.1 release (standalone backport for users on stable release)

Testing: Verified Gemma3-12B LoRA on ChartQA - training proceeds past initialization with warning log