Use sparsified neural models for relevance task

[Shreyanand]

The existing transformer model use distilbert but we want to see if there are other models in the sparse zoo that we can train for a better performance. Use the existing notebooks (transformer_relevance, transformer_inference) as a reference and create new benchmarks for this task.

Acceptance

• Replicate the training and inference snippets used for the distilbert model to the sparse model
• Get the benchmarks for the sparse model (note: add model size comparison to benchmarks)
• Explore other models in the sparse zoo and select a few to add to the benchmarks