Add single-precision kernel for calcMahalDistGpu

We have substantially simplified the kernel, which has made it faster
and able to handle larger D. We now use this kernel for D <= 64.

For larger D, we have switched from a triangular solve (TRSM) to
computing the explicit inverse (TRTRI) and performing a general matrix
multiplication (GEMM). This is because TRSM is much slower than GEMM
on the GPU, despite the increased number of operations. For D > 2048,
we stick with TRSM.

Also, add -lmwlapack option to buildMexFiles, since calcMahalDistGpu
needs to be linked to LAPACK for the triangular inverse.

@MoDT/buildMexFiles.m

@@ -52,6 +52,7 @@ function buildMexFiles()
 % Compiler/linker options
 mexcuda_opts = {
     '-lcublas'                      % Link to cuBLAS
+    '-lmwlapack'                    % Link to LAPACK
     ['NVCCFLAGS="' nvcc_opts '"']
     ['CXXFLAGS="--compiler-options=' compile_opts '"']
     '-L/usr/local/cuda/lib64'       % Location of CUDA libraries