Improved Patch Boundary Handling In LSSTDataSet cutouts

@mtauraso is already working on this.

Since Apr 29, 2025.

Currently, via https://github.com/lincc-frameworks/hyrax/pull/304, we error out if the size of the cutout exceed the patch size. Pasting discussion from Slack here:-

**Aritra**
1. I like this as a first draft, but of course, it’s limiting because for science efforts that use larger cutouts; this will fail a lot.
2. Protected under a flag and argument; this is a great idea. You should let the user choose either the function to error out; or pad with a user-specified value. Different users will have strong preferences on whether to pad with nans, 0s, or the (weird astronomical choice of) -99s.

**Michael**
We have a general need for nan-filling from SSC. I think we want a global config for nan handling, but haven't thought through all the cases yet. My understanding is that for nan-infilling you generally pick an algorithm not a static value. If I remember colin's explanation right, for most calibrated cutouts there isn't a value that approximates "nothing there" due to noise+calibration, which is why you do something like "fill with 5th percentile value of the image" or the like, to approximate what would have happened had the telescope collected data in the area of nans

**Aritra**
I think different things will work well for deterministic vs ML algorithms. I think I have often found in imputation kind of situations, you need fixed values for CNNs. Either way, it seems like we will ned to come up with some flexible configs to help users come up with solutions.