Fix indexing errors

[austinpeel]

No description provided.

[aymgal]

I notice that the build is failing because of the spectral norm test. I don't know if this is only a numerical issue, but the difference with the previous value is quite significant. Just to be sure, did you check that your fixed version gives the same result than before, on simple test cases (excluding extreme lens models) ?

[aymgal]

Awesome, thanks! I gave you access to merge, so when you're done with this PR please merge it ;)

[austinpeel]

Will do!
So yeah, what was happening was that in the case of source pixels on the border whose neighbors were out of bounds, the weights were being calculated as if those neighbors were still included. So when those neighbors were excluded at the end, the sum of the actual weights was < 1 in those cases. To fix it, I just scaled up those weights to properly sum to 1, and this in turn fixed the spectral norm issue.