Fixed bug in building convolution windows for mrCOSTS

[klapo]

Hi all,

I found a small bug in how mrCOSTS builds it convolution windows and performed window rounding. The long story short was that I didn't correctly adapt the original MATLAB indexing for python. This pull request fixes that issue.

Summary of changes:

• Convolution windows had MATLAB style indexing and not python. As a result the windows did not peak at the right place.
• Similarly, the window rounding was asymmetric.
• Causes slight reduction in the reconstructed error in the toy data example.
• b amplitudes now fall along PSD of the input data, fixing a lingering issue in the mrCOSTS fit and giving it more diagnostic power.
• Fixed the mrCOSTS tests not correctly tearing down.

Thanks for taking a look! Karl