Implement Min-P sampling and late temperature adjustment as a fused sampling layer

[aikitoria]

This is a work in progress implementation of #1154 implementing both the Min-P sampling and the option to compute temperature after the Min-P filter.

It first determines the maximum P element, then computes the sum of adjusted probabilities over all unfiltered elements, and then directly samples one in the range.

What makes this a draft:

• It cannot be controlled from the frontend with a separate min_p parameter. But how could I implement this, when some random auxiliary code (executor) is closed source for no apparent reason?
• It currently does not support processing requests that want top_k/top_p in the same batch as ones that want min_p. I am of the opinion that enabling top_k/top_p and min_p on the same request is nonsense.
• It does not yet support exporting logprobs if requested by the user.
• There are no tests.