Updating bfloat16

• Change Nvidia apex to the native pytorch
• Support bfloat16 (by --bf16)
• Fix the USAM error
• Adding DINO style loss and EMA for --DG