Extract eh_proj Layer from ParallelLMHead for MTP to Avoid Weight Transposition Issue

[Deleter-D]

When using MTP and tie_word_embeddings=true, unexpected weight transposition behavior may occur.

The specific reason is that the eh_proj layer in MTP reuses the ParallelLMHead class. When MTP is turned on, its internal judgment logic about tie_word_embeddings will cause the eh_proj weight to produce unexpected transposition behavior. Therefore, the eh_proj layer of MTP is extracted separately to facilitate subsequent development.

[paddle-bot]

Thanks for your contribution!

[CLAassistant]

All committers have signed the CLA.