sampling : avoid expensive softmax during greedy sampling

[ggerganov]

fix #9530

When the temperature is non-positive, we can simply sample greedily the token with the highest logit. But in some cases, the probs of the secondary tokens are also required (e.g. llama-server to display candidate probs, llama-speculative to peform stochastic speculative sampling). In such cases, we first filter the the top sparams.n_probs tokens via a top-k sampler and then apply softmax to them in order to avoid sorting the full vocabulary.

Also add perf timings to test-sampling to keep track of the performance of the samplers.

• I have read the contributing guidelines
• Self-reported review complexity:
  • Low
  • Medium
  • High
    ggml-ci