Added optional GPU Memory Logging

[krishgera]

This pull request introduces an optional --log-gpu-memory command-line flag in generate.py to log GPU memory usage at key points (immediately after loading the model, before generation, and after generation). By default, logging is disabled, so there is no impact on users who do not require memory tracking.

Changes :
Added a --log-gpu-memory argument.
Conditionally log GPU memory allocated/reserved at each relevant inference stage.

Rationale :
Simplifies troubleshooting for users running large models on limited GPU VRAM.
Maintains existing code paths when the flag is not used.

Testing :
Verified correct parsing of the new flag in local setups.
Observed expected memory logs appear only when --log-gpu-memory is enabled.

No additional dependencies or performance overhead for users who opt out of memory logging.

[Some1and2-XC]

You should remove the trailing indentation.

[krishgera]

You should remove the trailing indentation.

Fixed trailing indentation. Let me know if anything else needs adjustment.

[AlphaNewt7]

An interesting inclusion.

[krishgera]

@mowentian can you review this?