Add PaLM-2 models [EXPERIMENTAL]

[JosselinSomervilleRoberts]

This adds 5 PaLM-2 models and a VertexAI client:

• text-bison@001
• text-bison-32k
• text-unicorn@001
• code-bison@001
• code-bison-32k
  Other models are available here but I only added those necessary and relevant to HELM (although Imagen could be added for HEIM).

There are several issues related to the lack of documentation that have been opened:

• Add PaLM-2 tokenizer when available #2083, no tokenizer is provided so we use google/t5-11b which is probably not correct.
• Support more parameters for PaLM-2 #2084, some fields of Request are not supported because the Python SDK is incomplete.
• Support logprobs for PaLM-2 models #2085, some fields of RequestResponse are not supported (such as logprobs) because the Python SDK is incomplete.

I have not been able to verify the context length due to the very small quota I have and used what is provided in the API ref.
We should add a max_output_tokens to WindowService as some models have a very low value for this which would probably break on some scenarios (code-gecko@001 had a limit of 64 output tokens). (See #2086)