Support for gradient accumulation #9

[fluoryyn-art]

1. Added gradient accumulation parameter to argument.py with default value = 1 (maintain backward compatibility)

2. Updated training function in utils.py to handle gradient accumulation by:

   • scale the loss by 1/gradient_accumulation_steps to maintain same effective learning rate
   • only perform optimizer steps after accumulating gradients from the specified number of steps
   • handle final steps to ensure gradients are applied even if steps don't divide evenly

3. modified run.py to update deepspeed configuration with gradient accumulation steps

[fluoryyn-art]

#9