[BUG][Performance] matrix::detail::select::radix: inefficient "one-block" kernel launch

[achirkin]

matrix::detail::select::radix contains two almost identical implementations of the radix MSD select algorithm. These are radix_kernel and radix_topk_one_block_kernel. The main difference is that the one-block kernel is tailored for somewhat larger batch sizes: it runs only one block per row, and thus does not require any inter-block communication. There is, however a two-fold problem with the one-block kernel:

1. It uses the same function calc_chunk_size() to select the CUDA grid size as the normal kernel; as a result, sometimes very small grid sizes are selected and the algorithm runs at a very low occupancy in an inefficient loop over the input batch.
2. It allocates the temporary buffers of the size input_row_length * gridDim * 2, which can become extremely large and inefficient if we fix (1). In contrast, the normal kernel has an optimization to use a limited-size temporary buffers.

By default, this problem is masked by the matrix::select_k heuristic. It just selects a faster legacy implementation borrowed from FAISS for the problematic input sizes.