Provide BiCGSTAB with no extra dependency

[tvercaut]

This is a follow-up from #3

Options include adapting code from:

• PyKrilov, not in pytorch by pure numpy so easily adapted https://github.com/PythonOptimizers/pykrylov/blob/master/pykrylov/bicgstab/bicgstab.py
• https://gist.github.com/bridgesign/f421f69ad4a3858430e5e235bccde8c6 (vanilla implementation of the version on wikipedia -> may not be robust enough)