Standardize Input Size Parameter and Add Progressive Training Features

[dnth]

This PR includes several improvements to parameter naming and training capabilities:

Key Changes:

• Renamed --video-width parameter to --inference-size in live inference scripts for better clarity and consistency
• Added progressive resizing training strategy with notebook demonstration
• Updated training script configurations for improved batch handling
• Enhanced checkpoint loading with better error handling for image size mismatches

The parameter renaming provides more intuitive configuration options for users, while the progressive resizing strategy allows for more efficient training workflows. The checkpoint loading improvements make the system more robust when handling models trained with different configurations.

Testing:

• Verified parameter changes in live inference scripts
• Tested progressive resizing strategy with 128px initial size
• Validated checkpoint loading with mismatched configurations