79 improve pairwise registration

[FIrgolitsch]

I've added a rigid registration step before the affine registration, only if the affine is chosen. Both transforms are then placed in a composite transform and saved to disk.
Other changes:

• Reduced the max iterations to 1500
• Reduced the gradient magnitude tolerance to 1e-6 by default. I looked around a little, and it seems for noise images the recommended value is 1e-6 to 1e-8. As our images are quite noisy, I reduced the value. Quick testing showed it looked fine, and the convergence times are so much faster now (naturally). The register_pairwise step now only takes about 10 minutes.

Closes #79

@CHrlS98 Let me know what you think!

[FIrgolitsch]

Some quick images of the reconstruction with this:

[FIrgolitsch]

Next, I want to try adding masks. I think the registration process is still trying to register on the background tiles. The masks can also be used to clean up any background noise.

[FIrgolitsch]

I've added the masks. Here are some updated images of the reconstruction. The settings are exactly the same as the previous images, just with the new masks. I still need to verify the mask creation; it's a very simple Otsu threshold right now. As the maskless version seems better, I think it needs some work still.

[CHrlS98]

Also a comment on the masks. Since the normalization is done prior to registration, the background should be pretty close to 0 when registration occurs.

[FIrgolitsch]

@CHrlS98 I think it's ready. The channel logic was a bit of a mess, but I got there. I also changed all the copy directives for the output files to move. I don't think we should duplicate some of that data, as it can quickly get out of hand.