feat: use a KDE for smooth scores over tissue area

[jsicherman]

Performs both a weighted (by score) KDE and an unweighted KDE on the cell coordinates, and computes their ratio across a grid of n_bins x n_bins. This results in much smoother score estimates across the tissue. Addresses #7

Remaining work:

• Implement a way to evaluate the estimator at cell coordinates/neighborhoods
• Convert to smooth image mask for region exclusion (could also be done in a follow on PR)... Are there geometric properties of these masks which could be used to estimate something about the type of failure?

Binned:

Using estimator:

Different dataset, binned:

Using estimator:

Sampled atop cells:

As a mask (thresholding at the 70th percentile):

[berl]

transcribing from our slack conversation:
pretty sure KDE on its own was the wrong idea. However, it seems like a weighted KDE divided by a non-weighted KDE would give a “kernel-mean-estimator” which is what I was envisioning: a kernel weighted mean value of nearby expression ratios

there is probably a name and package for this. note that the scikit-learn KDE has a ball tree under the hood, so (edit: speed) comparison of this to a nearest-neighbors approach may be a wash. one nice thing about the sklearn KDE is that they have few different kernels that can speed up computation

[berl]

I'm excited to get this in! I added back the fixed range for the z-score plot. comparing the KDE to the binned version shows some differences that I'll add another issue for figuring this out: