Use half2 and float2 types/operators for Vamana build L2 metrics

The Vamana index build has several different L2 comparator methods based on the dataset dimension and datatype. For large dimensions, each thread loads several values and accumulates the total to increase ILP, but each value is loaded independently. For the __half and float types we can improve performance by using the half2 and float2 types and using the __hfma2, __hsub2, __fmaf2, and __fsub2 operators to calculate and accumulate the L2 distnace.

Start by locating the Vamana index-build L2 comparator methods for large dimensions and the __half and float paths. Compare how values are currently loaded and accumulated, then verify that the relevant paths use half2 or float2 with the named operators and preserve correct L2 distance results.