copy flat data maybe faster

[wiyr]

https://github.com/microsoft/DeepSpeed/blob/master/deepspeed/runtime/fp16/fused_optimizer.py#L298C1-L301C37

Since fp16_groups shared memory with fp16_groups_flat, fp16_groups_flat format is the same as fp32_groups_flat，so we can just copy fp32_groups_flat to fp16_groups_flat without unflatten ?