Adding data conversion and training script to support HF Accelerate

[nkatyal]

The following changes introduces a newer version of the training script that uses HF accelerate. The functionality remains the same while making it easier for the user to set up the training environment. The changes has been tested in the following environments

🔍 Model Training Comparison (Old vs. New Setup)

Model	GPUs	Training Time (Old / New)	MRR (Old / New)
BERT Short	1 × NVIDIA L4 (24GB)	3h 30m / 3h 22m	0.3319 / 0.3401
ElectraBERT MAXP	4 × NVIDIA T4 (16GB each)	6h 15m / 6h 30m	0.3364 / 0.3428