Fix bug with number of evaluation steps (EleutherAI#384)

we were running way to many evaluation steps if the model is pipe parallel + has g.a.s on because of this line

```python
            for _ in range(neox_args.gradient_accumulation_steps):
```

- fixing this to 1 if the model is pipe parallel fixes the issue, as .eval_batch() already takes gradient accumulation steps into account.

megatron/training.py

@@ -528,7 +528,8 @@ def evaluate(neox_args, forward_step_fn, data_iterator, model, verbose=False, ti
 
  # although we're not accumulating gradients here, we count one iter as train_batch_size_per_gpu * g.a.s
  # to be consistent with deepspeed's pipe parallel engine
- for _ in range(neox_args.gradient_accumulation_steps):
+ # since pipe parallel already takes gas into account - default to 1 here if pipe parallel is true
+ for _ in range(1 if neox_args.is_pipe_parallel else neox_args.gradient_accumulation_steps):
  # Forward evaluation
  loss = forward_step_fn(model=model, data_iterator=data_iterator, neox_args=neox_args, timers=timers)
  losses.append(loss)