Fix scheduler stepping and label dtype handling in training loop

[Apprentice2907]

Fixes #152

Changes Made

1. Scheduler Fix

   • Replaced scheduler.step(loss) with scheduler.step()
   • CosineAnnealingWarmRestarts does not accept a loss argument and expects a simple step call

2. Label dtype Fix

   • Replaced labels.type(torch.LongTensor).to(device) with labels.long().to(device)
   • The old approach creates a CPU tensor first before moving to device, causing unnecessary CPU→GPU transfer
   • The new approach keeps the tensor on the correct device throughout

Files Changed

• DeepLense_Classification_Transformers_Archil_Srivastava/train.py