Implement new sampler API and bump llama.cpp

[volesen]

llama.cpp introduces a new sampling API in b3680. This PR aims to replace the previous custom sampling API, with a wrapper around the new llama.cpp sampling abstraction.

We remove the custom sampling functions from LlamaContext, so this is a breaking change.

Furthermore, we bump llama.cpp to b3750, which finalizes the new llama_perf features, so we can more easily preserve the timings API.

Credits to @brittlewis12, whose draft we copied in large parts. (#548)

It's probably worth it to squash commits before merging.

The new sampler is showcased in the simple example and can be used like this:

let sampler_params = LlamaSamplerChainParams::default();
let mut sampler = LlamaSampler::new(sampler_params)?.add_dist(seed.unwrap_or(1234));
let new_token = sampler.sample(&ctx, batch.n_tokens() - 1);
sampler.accept(token);

[MarcusDunn]

awesome work! I'll get this in + release soon as checks look good.