added debugging to train.py

[Subhanshusethi]

1. Added a valid indices check which, while loading the tokens file, ensures that the number of captions matches the number of embeddings. If not, the mismatched entries are filtered out.
2. Fine-tuning GPT-2 is a major task that highly relies on the dataset. Removed single letters, special characters, and stop words (as defined by NLTK by default) to reduce the impact of connector words while training in the embedding space.