Enable T2U Finetuning for UnitYNART2UModel in SeamlessM4T_v2_large and Refactor Context Management

[Nabarup1]

This pull request introduces support for finetuning the T2U (Text-to-Unit) component in SeamlessM4T_v2_large models that utilize the UnitYNART2UModel, which was previously not supported in the official finetuning trainer.py. The UnitYFinetuneWrapper has been extended to detect and handle both UnitYT2UModel (v1) and UnitYNART2UModel (v2), enabling comprehensive T2U finetuning for v2 models. Additionally, the dummy_context has been replaced with nullcontext() to ensure robust context management and prevent attribute errors during training. These changes ensure that users can now reliably finetune the T2U module in SeamlessM4T_v2_large without encountering runtime errors or missing model support.

What has been done:

• Added explicit support for UnitYNART2UModel in the T2U finetuning workflow.
• Updated the forward logic to accommodate the specific interface and output shape of UnitYNART2UModel.
• Replaced the use of dummy_context with nullcontext() throughout the trainer to prevent context-related attribute errors.
• Improved error messaging for unsupported T2U model types, clarifying which models are supported for finetuning[^1].

[Nabarup1]

@samueljamesbell @Celebio @mhlakhani
Kindly requesting your review on this PR, which introduces T2U finetuning support for UnitYNART2UModel (SeamlessM4T_v2_large) in trainer.py and updates context handling for improved reliability. Your insights and feedback would be highly valuable.

[Omaralsaabi]

Thank you, that saved me a lot of time, but I made a change in line 334 to make it work.
find_unused = self.params.finetune_mode in [FinetuneMode.TEXT_TO_SPEECH, FinetuneMode.SPEECH_TO_SPEECH]

[Nabarup1]

Appreciate you testing this out, @Omaralsaabi — thanks for pointing out the find_unused tweak. Makes sense. I can fold that into this PR so others don’t run into the same issue.

[Nabarup1]

Hey @samueljamesbell @Celebio @mhlakhani — just circling back on this. Checks are all green, and with @Omaralsaabi's feedback addressed it should be good to go. Would be great to get a review/approval when you get a chance 🙏