Update TGI specs (+ define TextToSpeech independently)

[Wauplin]

This PR pulls latest changes from TGI specs. In particular:

• updated ChatCompletionInputMessageContent => now supports single message (as before), array of messages and sending image as input (in case of VLMs)
• new input: stream_options
• tools input have been updated
• more documented fields

Note: TGI specs have a guideline input parameter for chat completion. I removed it from the official specs as it's a very specific parameter and subject to deletion in the near future (it was there to fix a chat template issue for ShieldGemma models). More context here (internal).

Note: I also took the opportunity to define TextToSpeech specs independently from TextToAudio (otherwise the semi-automatic scripts are not working properly). Nothing has changed in the specs.

I'll open a related PR on huggingface_hub side to reflect this change.

[julien-c]

general comment that this would really need to be maintained by TGI's maintainers IMO... poke @OlivierDehaene @ErikKaum @Narsil

[Wauplin]

Yes, or running it in a CI since it's just a script to run now

[mishig25]

from my side, lgtm !
checked the generated inference.ts interfaces

[Wauplin]

Thanks for checking the JS part @mishig25! I've used and checked the Python part myself so let's say it's fine to merge.