[ML] Sanity checks for properties of trained boosted trees

[valeriy42]

• Check if the tree consists of a single stub. Produce a warning if this is true.
• Check that tree doesn't predict NAN. Report an error if a NAN is detected.
• Check if a recall is 0 on one of the classes. Report a warning if this is true.

Implementation considerations:

• This sanity check can be applied after predict() is called.
• The sanity checkers are specific to the used loss function. Hence, they can be acquired from the CLossimplementations.
• Instead of checks for recall, regression losses may have a stub method.