Fix: calculate Cooks distances with few samples

[BorisMuzellec]

This PR aims to fix the bug described in Issue #43.

The issue seemed to be that the summation was done on the wrong axis in the trimmed_variance function, called through robust_method_of_moments_disp in the calculate_cooks method of the DeseqDataSet class.

[ghost]

Very high level review, but is there any way to add a unit test to check the behaviour that lead to this bug? This way, we make sure the codebase does not forget about it. Also, it's a good way to ensure this PR really fixes the related issue.

[BorisMuzellec]

Very high level review, but is there any way to add a unit test to check the behaviour that lead to this bug?

I guess we could add a test where the pipeline is run on a dataset with few samples, as the piece of code responsible for the bug only is run only when no cohort has 3 or more samples. Not sure what to test for though, would error-free execution be enough?

[ghost]

Not sure what to test for though, would error-free execution be enough?

Yes I think that simply adding this corner case and ensuring it runs without error is a good starting point!

[ghost]

LGTM, thanks @BorisMuzellec