automatic step size reduction code a little.

scripts so they attempt to verify the version of cuDNN you have installed and
print a useful message if that isn't the case.

orthogonalize does not work with fixed-sized matrices

detect OS X, and query sysconfig for libdir

minor things.

once every few thousand iterations.

global functions.

inside mex files.  This way, if you use matrix_colmajor or fmatrix_colmajor
in a mex file it will not do any unnecessary copying or transposing.

ODR violations.

and LAPACK that comes with MATLAB since trying to use any other BLAS or LAPACK
generally makes MATLAB unstable.