Support long responses and additional fixes

[mayabar]

Currently, responses are generated by randomly selecting one sentence from a predefined list. As a result, even when max_token is set to a large value, long messages are never returned.

The suggestion to use a lorem ipsum generator was declined because the library we identified can produce only up to 191 words; any request for a larger word count causes it to panic.

New behavior:

• Echo mode: The input text is returned as-is, if max_tokens or max_completion_tokens is set to value higher than the input tokens number, if max_tokens or max_completion_tokens is lower - the input will be trimmed. Useful for testing where we need to know exact response in advance.
• Random mode:
  • if max_tokens or max_completion_tokens is specified, sentences are selected from the predefined collection until the token count reaches the limit
  • if number of output tokens is not defined in the request, a random token count is generated for the response. This count is based on a Gaussian distribution with mean of 40 and standard deviation of 20, capped by a maximum response length which is currently set to 128 tokens.

Additional changes:

• Use tokenize function which divide text by space and additional characters in request processing too (not only in tools related part)
• Validate max_token and max_completion_token as request arrived and return 400 status if it's invalid
• Protect generating any random value with mutex
• Fix test for the changes above + add test for random texts creation

[shmuelk]

/lgtm

/approve