Evaluation Update

[SHUMKASHUN]

Fix some style problems in PR351.

[research4pan]

LGTM, thanks! Some further improvement tricks:

src/lmflow/pipeline/evaluator.py

• If most code of _evaluate_acc_with_accelerator and _evaluate_acc_with_deepspeed are the same, write them together would be a better option, since modifying the accuracy computation logics once can result in support of both modes.
• [Style] line 224: if want a detailed logging message format, using logger.info instead of print would be better.
• [Style] line 226: if (self.evaluator_args.use_wandb == True) -> if self.evaluator_args.use_wandb.
• [Feature] Better raise error for other metrics ppl and nll if accelerator option is specified but not supported.