Do not update past_key_values in place

[irenedea]

Fixes the forward pass in MPTModel to update not past_key_values in place and instead use a new list, presents. 🎁 Naming is inspired by gpt2.

This addresses the following error:

 RuntimeError: attn_bias (shape: torch.Size([1, 4, 1, 2])) is expected to broadcast to shape: torch.Size([1, 4, 1, 3]).

which occurs during multi-gpu generation for mpt models when one rank hits EOS before the others.

[dakinggg]

LGTM apart from my comment on the test. Also wait for @vchiley to look please.

[vchiley]

nit: presents implies the current x, not the past...
the tensor its referring to is the past and presents.
no clue why gpt2 names it like this.
this is a nit because its just a naming convention comment.

otherwise lgtm

[dakinggg]

@irenedea is this ready to merge?

[dakinggg]

one nit