Description

The related issue: #240
In this PR, a new model is added: semantic id based decoder.
Some highlights:

1. We use mcore standard transformer block as the backbone of our decoder.However, our jagged input and beam search require special masks, as a result, we have to create padded dense attention mask and invoke the local impl of SPDA of mcore.
2. We implemented a simple beam search module, which is called during evaluation.
3. The model is not SOTA impl, and the convergence needs to be confirmed. Only amazon beauty dataset is tested.
4. We now only support single GPU

Checklist

• Code Cleaning
• README
• CI pipeline ( integration test)

CI:
CI