optimize GPU execution and eliminate CPU sync bottlenecks

[BehradFM]

Hi there

I was training this model on my GPU and noticed some performance bottlenecks caused by frequent CPU-GPU synchronizations.
I've refactored parts of the code to keep tensors on the device and maximize hardware utilization.

These optimizations resulted in a massive performance boost, making the overall training process approximately 2.5x faster!

• Batching Optimization: Move the identity matrix ("I") used for batching directly to the device (CPU,CUDA).
• Native Operations: Rewrote "choice_accuracy" function.
• Asynchronous Execution: "Disabled CUDA_LAUNCH_BLOCKING" to allow PyTorch to execute operations asynchronously.

I have also verified that the code remains fully compatible and works on both CPU and GPU device.

I would be really happy to see this merged so that others in the community can benefit from the faster training times!
Thanks for the great repository!

[BehradFM]

• I have just recently Refactored l1_regularization function to dynamically initialize the accumulator tensor on the target device.