[ML] Apply source query on data frame analytics memory estimation

[dimitris-athanasiou]

This is the same fix as #49517 but as the code around the fix has changed
quite a bit, the fix had to be reworked. The idea is the same though.

Closes #49454

[elasticmachine]

Pinging @elastic/ml-core (:ml)

[przemekwitek]

LGTM

[dimitris-athanasiou]

Holding off merging this in until 7.5.0 goes out.

[benwtrent]

@dimitris-athanasiou does this change also take into account different training percentages? As those should effect memory estimations if they do not.

[dimitris-athanasiou]

@benwtrent Regardless of the value of training_percent we load both training and non-training rows into the c++ data frame. We do so as predictions are calculated for non-training rows. Thus, I don't think this affects the memory estimation.