Restored masked scaled dot attention

[l-k-11235]

Revert changes inserted by #2538 and adapt to #2539.

[vince62s]

For the sake of clarity:
when using left-padding at inference with LM models, we need a mask even for step > 0 in order to mask the padded positions on the left hand side of the batch. However pytorch SDPA does not accept a mask for non causal, hence we need to use this mask only for the "regular" scaled dot product attention manual calculation.
merging.