chore: make vector validation work with fp16

[chebbyChefNEQ]

two issues fixed:

• when using fp16 and cosine together, the distance we get is usually greater than 1e-6 even when we hit an exact match.
  • fixed by comparing the vectors instead of distance
• fp16 can have edge cases where IVF partition is assigned wrong (very low percentage, fixable by just using more nprobes)
  • fixed by allowing lower validation threshold for fp16