common class for aligner models

[gshennvm]

Is your feature request related to a problem? Please describe.

Many different classes in aligner have similar implementations that are duplicated. For example log prob calculation in DPO and PPO get log prob. Instead we should create a common class or common until function to avoid this duplication.

in complicated cases like with TRT-LLM this is even more important since there are various edge cases that should be handled centrally.