Variable DEQUANTIZED_MIN_ROWS`Const`

DEQUANTIZED_MIN_ROWS: 128

The fewest rows the dense GEMM takes at any projection: below this the dequantization costs about what the GEMM saves over MLX's quantized matmul.