Osimons/gpu sampling equivalence checks

[ORippler]

Spun up a script to investigate differences between backend-samplers and samplers working inside llama-cli/outside of ggml's opset. Main investigation insights:

1. std::uniform_real_distribution<double> returns different vals than std::uniform_real_distribution<float> when seeded with the same rng.
2. Apart from top k not having to return sorted outputs, main differences come from soft_max and cumsum, which will behave differently in ggml than in top-p/dist-sampler's due being parallelizeable across cores/sm
3. Tried changing dist-sampler formulation (threshold based on cumsum(exp(logits)) vs. threshold based on (cumsum(probs)), did not yield that many differences. Thus did not push the code
4. We use unstable std::partial_sort in llama-cpp samplers for sorting

After this PR + fixing warmup currently advancing rng for backend-samplers by 1 call, outputs are much closer for backend-based compared to llama-cpp-based sampling.

Filed this as a separate PR as I'm not sure we want the vibe-coded script in main llama.cpp