lgmres for numpy backend

[mganahl]

[alewis]

It would be super cool if on CPU (or maybe just in NumPy) we could call this code, which is a drop-in replacement to the SciPy version that is apparently 30 times faster: https://github.com/haranjackson/LGMRES. This seems to be possible under the MIT license that it's under, but I'm not sure if it's ok re: Google - we would need to ship the code, and I guess it would need to be compiled.

@Thenerdstation

[alewis]

Upon a closer read the code in question only accepts dense matrices (which is weird because the C++ code is actually in terms of a matvec function). So that pretty much sinks that dream.

[mganahl]

well, we can always write our own extension if we want