FIX switch to binary_crossentropy if a binary problem

[adrinjalali]

Fixes #14858.

Raise a warning and switch to binary_crossentropy if it's a binary classification problem and categorical_crossentropy is given.

[NicolasHug]

Thanks for the PR.

If the user specifies categorical_crossentropy on a binary classif problem, I think we should either:

1. raise an error
2. still use categorical_crossentropy but raise a warning "you're building 2 trees instead of 1 so just use 'binary_crossentropy' or even better, use 'auto' ".

I'm in favor of just raising an error (because we have the 'auto' option that just works). But I'm OK with option 2 as well. The current proposal is re-implementing the 'auto' logic but with a warning so I don't think we want that.

[NicolasHug]

Thanks!

[adrinjalali]

@ogrisel maybe you could have a look?

[adrinjalali]

ping @ogrisel or @glemaitre should be an easy one :)

[glemaitre]

Apart of the comment raised by @thomasjpfan LGTM

[glemaitre]

I pushed the fix and will merge when it is green