Fix /vertex payload parsing when MESSAGES_API_ENABLED

[alvarobartt]

What does this PR do?

This PR fixes the input payload parsing on Vertex AI i.e. when compiled with the google feature and deployed in a Vertex AI environment (replicable with the following environment variables AIP_MODE=PREDICTION, AIP_HTTP_PORT=80, AIP_PREDICT_ROUTE=/predict, and AIP_HEALTH_ROUTE=/health); when the environment variable MESSAGES_API_ENABLED is set to true, meaning that the /vertex endpoint is a proxy for /v1/chat/completions endpoint instead of the default /generate.

Note

As already discussed with @Narsil this is most likely not the best approach, here are some things that I tried:

• Implementing the From trait to go from the input Vertex AI payload into ChatRequest, but that still requires us to have a "duplicated" struct holding the generation kwargs, which is not ideal
• The above could be solved by moving the generation kwargs out of ChatRequest whilst just leaving messages there, and using #[serde(flatten)] over the generation kwargs; but that's not ideal either
  Any feedback is appreciated!

Who can review?

@Narsil as of our previous conversation today, feel free to take over the PR, or even close it in favour of a potentially better PR! Thanks in advance Nicolas! 🤗

[alvarobartt]

Closed in favour of #2553, thank you @Narsil!