Handling of missings

Not sure if it's the appropriate place to post this; it looks like we could improve the situations when dealing with missings. Here's a simple example of something that's not ideal:

1. y has some missings
2. X doesn't
3. we drop the rows of y and X that have missings
4. we feed that in a machine
5. the machine fails because the model doesn't handle Union{Missing,Float64}

of course there are no missings in y anymore. Interestingly, the scientific type is correct so that could be exploited but the machine type has Union{Missing,...}

Maybe we could help with:

• provide a dropmissing function which "re-coerces"
• have a mechanism in fit which checks whether the type and scientific type match in terms of missing, if they don't and ST has it right, maybe the data can be re-coerced

thoughts?

Start by reviewing MLJBase's existing handling of missing values, scientific types, and the fit path described in the issue. The scope is not settled: compare the proposed dropmissing re-coercion with fit-time checking, then clarify the intended behavior and add tests demonstrating that cleaned data no longer retains an incompatible missing-value type.