Anthropic max_tokens is 4096 by default

[DouweM]

As of #1994, the default for Anthropic max_tokens is 4096. I had previously made it use the models' actual max tokens limits by default in #1979, but that was hitting "a ValueError if a non-streaming request is expected to be above roughly 10 minutes long": https://github.com/anthropics/anthropic-sdk-python#long-requests

I don't like this though as it's pretty surprising to users, I'd rather have it intelligently use the highest number that fits under the ValueError limit for non-streaming requests, and the actual token limit for streaming requests.

Note that we'd need something like #2067 to not hit this error on Vertex AI:

{"type":"error","error":{"type":"invalid_request_error","message":"input length and `max_tokens` exceed context limit: 189127 + 16000 > 200000, decrease
input length or `max_tokens` and try again"}