[BUG] SVDBPI nans emitted

[Jonathan Wong]

If one supplies a complex matrix A of say {16,16} which contains several columns of zeros, svdbpi will return nan values in U, S, VT. However, if one uses a truncated A which contains less columns of zeros (e.g. {16x4}), no Nans are emitted.

To Reproduce
Patch the included SVD.cu matx test.

./test/matx_test --gtest_filter=*SVDBPI_bug

Expected Behavior
Zeros should be emitted instead of Nans.

Did not test this with non-complex types.
SVD_bug.tar.gz