Implement Multi Token Prediction loss in Qwen3.5

[tomiock]

use the already existing MTP module of the model to implement the loss during training via next token prediction