Suggestion for basic support:
expose blending_frame param.
if blending enabled, we don't need to force keep_cross_validation_predictions to true when training base models.
leave CV default as is?
train the 2 SEs in blending mode instead of cv-stacking mode.
Alternative is to just expose a blending params:
automatically split the training_frame to generate a blending_frame (75/25).
disable CV (set nfolds to 0) for base models.
disable all keep_cross_validation_* params.
train the 2 SEs with generated blending frame.