Magma expect column-major matrices. Exchange u_data -> VT and vt_data -> U to
match numpy svd interface.

Added an optimization for inplace gpu_cusolver_solve and tests

Add on_error option for Cholesky

Fix more issues in DebugMode

Fix tests in debugmode

Fixed minor spelling error

Corrected some comments and code formatting

[Interface change] be more consistent, use ndim instead of outdim

Remove goto in OpenMP Elemwise fail code

avoid creating float64 variables

The regular python2 and python2_debug buildbots used the same cache directory.

Stop the compilation when we found a negative shape.