[FEA] Improved support for building graph-based indexes on lower dimensional vectors

While CAGRA, DIskANN and other index types tend to build indexes much faster for higher dimensional vectors on the GPU (e.g 10x+ speedup), we are finding that compressed, heavily quantized, and lower dimensional vectors (eg. 64 dims) are not as competitive and are yielding build speedups that are just not acceptable in practice (e.g. 4x or less) to convince users to invest in the complexity of GPUs.

The working theory is that the lower dimensions aren't able to fully saturate the memory bandwidth and cause low GPU utililization in general. This is also in part because the each vector tends to use random io, and so when each vector is not large enough to saturate the coalescing bandwidth, we leave a lot of perf on the table.

There's an assumption here that we can update the grids in the CUDA kernels to do more vectorized loads, or chunking within each warp so that we can improve the perf. We need to investigate the possiblities here because, while LLMs and RAG have typically driven the vector widths up, we are beginning to see a general shift towards compression, quantization, and smaller dims as a means to 1) scale, and 2) save on disk/memory costs.

No files or tests are named. Start by profiling CAGRA, DiskANN, and other graph-index build paths for lower-dimensional vectors, then inspect the CUDA kernels and their grid, load, and warp behavior. Done means demonstrating improved build performance for compressed or roughly 64-dimensional vectors, but the issue does not define a target speedup.