Add bfp16 support to inhibit transforming to fp32 when using llama3

[StarConnor]

For Llama3, it is trained on bfp16, so we should not transform bfp16 to fp32 during inference. So I change the two precision transformations in transformer_lens. Basically, I add torch.bfloat16 in the precision check list to avoid transformation to fp32.

1. Attention
   In abstract_attention.py.
2. Layer Norm
   In rms_norm.py, and rms_norm.py (because Llama3 use RMSnorm)
   Note: I did not change the precision check list in layernorm_pre.py and layernorm.py because Llama3 doesn't use them. Should I also modified these two files?

[dest1n1s]

From my perspective, changing other precision checks could be postponed to avoid other issues.