Feat: adding is_training argument to implement KV cache only during inference

[adityavipradas]

Resolving issue #92

• Added is_training argument in language_model.py and vision_language_model.py to populate KV cache only during inference and set it [None] during training.

[kashif]

if it helps, you should note that the nn.Module has a training variable so you can probably use self.training

[adityavipradas]

Makes sense. But I observed using self.training will result in populating KV cache during validation as well. Instead, I used torch.is_grad_enabled() which is set to False during validation and inference & True during training.

Created a new PR here: #94

[adityavipradas]

Closing this PR.