keep the minimum min_keep value to 1 in sampling

[kylo5aby]

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

Fixes #9587

[ggerganov]

I proposed an alternative fix in #9604. Think it solves the problem at a more fundamental level by fixing the min_keep logic in the tail-free sampler.