updated covariance calculation in Tractor, more robust to overlapping sources

[gemmahuai]

Previously, if any totally overlapping pairs exist in a given cutout, we would lose information for all sources being photometered simultaneously in the cutout even if they are well resolved from the degenerate pair. This happens because a fisher information matrix (FIM) is constructed for each given cutout, which is later inverted to compute the covariance matrix. The matrix inversion fails if any models are dependent due to perfectly overlapping PSFs.

In this update, totally covarying sources are combined into a single model in the FIM, allowing the matrix inversion to proceed despite any degenerate models in a given cutout. These degenerate sources are being tracked using a overlap dictionary, so that we later assign NaNs to the their flux uncertainties. In this way, we are still able to extract unbiased photometry from other non-degenerate sources in the cutout.

For more detailed tests and validation, please refer to the slides I put together.