[sdmccabe]

Resolves #38. #57 introduced utilities for thresholding matrices in various ways, but they were largely not used. This incorporates them into the existing graph reconstructors where feasible.

There are a few reconstructors that do not threshold. For example, the minimum spanning tree does not threshold (because it is a minimum spanning tree). The one edge case here (which I did not touch) is convergent cross-mapping, which thresholds a matrix of p-values (i.e., a significance test). This could, in principle, be covered by the threshold utility function, but I don't think it's entirely appropriate to do so. (Giving the method the ability to set the significance level to get a fixed average degree seems to me more problematic than dropping low correlations to get a fixed average degree, but opinions may vary here.)

As a side effect of implementing this feature, the fit() method has been changed across all implementations (including the base reconstructor) to take arbitrary named keyword arguments through **kwargs. This allows us to pass threshold parameters freely, even to non-thresholding methods (where the parameters are ignored), which makes using this feature much easier in practice (see the change to netrdExplorer as an illustration of why).

As a side effect, this PR also adds tests for the thresholding functions and cleans up a number of the imports to netrd.utilities.

Finally, I'll note that some of the language used in the different implementations differs; e.g., matrix versus weights. Instead of addressing that in code review of this PR I would suggest raising an issue and PR-ing that separately.

[sdmccabe]

This is a big PR, so we should definitely talk things through and carefully examine what we're doing before merging this.

[tlarock]

As discussed in the meeting 2/13, @sdmccabe will later add warnings in some places, particularly when the thresholding is called with default parameters (since they should be set per-dataset).