Use tolerances for affine matrix differences to allow for floating point errors

[tsalo]

Summary

nilearn.image.concat_imgs appears to be overly conservative in how it compares affines to determine if they can be concatenated. Currently, we tell users to overwrite their affines manually, which is suboptimal. Instead, we could "look at the zooms, offsets and angles and set a threshold for each. The zooms had better be within eps of each other, same with the offsets. The angles we can calculate differences in degrees and go from there" (@mattcieslak's idea). If DWI files are compatible according to these thresholds, then we can create a modified version of each image with the affine from the first image, then pass them all into concat_imgs.

Stems from #225 (comment).

[pollaro]

@tsalo What do you think these tolerances should be?

I'm looking at some of my differences, and the degree differences are < 1 degree and the zooms and offsets are 1.6ish at most and usually much much smaller.

For scans that the centers were all copy and pasted to the same location, any idea what would cause this?

[tsalo]

I don't actually know. I think that might be a better question for @mattcieslak or @araikes.