Missing value codes, per dataset or per feature

[raamana]

User-defined missing value codes would be useful in certain domains, instead of lumping them all into np.NaN

It could be per feature (as different features may have different ranges) or per dataset, or per attribute (covariate do tend to be missing often!)

[raamana]

Suggested this in numpy for discussion as well: https://github.com/numpy/numpy/issues/14254