Very roughly, an evaluation around +0.50 may be +=, around +1.00 may be +/-, and around +2.00 (and larger) may be +-.

Details will depend on the engine giving the centipawn evaluation.

That will depend on the person commenting on the game, of course. It will also depend on the evaluation region this takes place at.

Dropping from +9.00 down to +4.00 may make no difference at all for the outcome of the game (white is still clearly winning after that), so it may not really merit a question mark in the eyes of many.

Dropping from +4.00 down to +1.00 may be a question mark for many, although some would say it is merely ?! for example when the resulting position is still technically winning for white anyway.

Dropping from +1.10 down to +0.10 is clearly a ? in most cases (except the engine evaluations are not accurate, for example in some endgames it does not properly understand), since it turns a probable win for white into a probable draw.

Does this give you a good general idea?