reduce size of egamma energy regression payloads

[slava77]

The last update in #17506 has more than doubled the size of payloads consumed by EGExtraInfoModifierFromDB in memory (from 22 MB to 53 MB).

Do we really need the GBRForestD here, can the float precision suffice?
Can some other "compactification" be done?

@lgray @rafaellopesdesa @bendavid

[cmsbuild]

A new Issue was created by @slava77 Slava Krutelyov.

@davidlange6, @Dr15Jones, @smuzaffar can you please review it and eventually sign/assign? Thanks.

cms-bot commands are listed here

[slava77]

@franzoni @arunhep please take a not from alca side as well.
Thanks.

[bendavid]

So the double precision is definitely needed during the training for the likelihood-based loss functions for numerical stability reasons. The possibility of converting back to single precision after the training is done has not so far been explored.

In principle possible and I guess it should work, though a little bit of code would be needed (new constructor for the single precision classes taking the double precision version as input).