Model does not train without FP16

[kipgparker]

Currently model does not train without FP16

if you set fp16["enabled"] = false in the "small.yml" config and run pretrain_gpt2.py it crashes with an error

  File "pretrain_gpt2.py", line 26, in <module>
    iteration = train(
  File "/home/aleph/repos/neox/gpt-neox/megatron/training.py", line 461, in train
    overflow_monitor.check(skipped_iter)  # check for repeated overflow
  File "/home/aleph/repos/neox/gpt-neox/megatron/utils.py", line 366, in check
    pretrain(neox_args=neox_args)
  File "/home/aleph/repos/neox/gpt-neox/megatron/training.py", line 96, in pretrain
    if self.optimizer.overflow and len(self.history) == self.n and all(self.history):
AttributeError: 'FusedAdam' object has no attribute 'overflow'```

Running on 2x3090's

[sdtblck]

@kipgparker that's intended behaviour for now - deepspeed in general does not work without fp16
My bf16 branch is almost ready to merge, though.