Anthropic - set max_tokens based on available context

[krissetto]

Calculate max_tokens value for Anthropic requests based on the available context using the official token counting API.

The PR fixes errors such as:

400 Bad Request

{
  "type": "error",
  "error": {
    "type": "invalid_request_error",
    "message": "input length and `max_tokens` exceed context limit: 178959 + 64000 > 200000, decrease input length or `max_tokens` and try again"
  },
  "request_id":"req_011CUB2XUHWkFnQzxcKBbjpF"
}

Closes #569

[krissetto]

Right now I implemented this calculation using the official Anthropic count tokens API.

Another option would be to do some approximated calculation with some external library, and use that estimate instead of an additional API call.

[rumpl]

Rather than doing double calls all the time, can we parse the error and do it only when needed?

[krissetto]

Rather than doing double calls all the time, can we parse the error and do it only when needed?

Sure, I'll take a look at that before merging

[dgageot]

@krissetto I'll merge it first. Please look at the improvement in a second PR