Write data loader for GPU training.

[afmjoaa]

We will use

1. nl-func
2. func-nl
3. func-func
4. nlf-nlf

As pre-trainer data and initial input token length for generating the next token will be a variable. This dynamic length need to be passed with each example.

[afmjoaa]

• [x] Task type isn't getting used in the category and task encoder, as the right value isn't passed properly

[afmjoaa]

• [x] Use response parser in inference

[afmjoaa]

• [x] Use the data loader in the training loop
• [x] Use the data loader in the inference loop

[afmjoaa]

Done