[BUG] a bug in get_instantiation_memory_required_per_gpu; memory cal error

There is a gpu memory calculation error in this function of
https://github.com/microsoft/DeepSpeed/blob/0a61d5d66462d51c6d867ab58479adf94f59de97/deepspeed/autotuning/autotuner.py#L278

        # assume the model uses Adam optimizer
        # ZeroStageEnum.disabled:
        params_mem = num_params * (2 if fp16_enabled else 4)
        gradients_mem = num_params * (2 if fp16_enabled else 4)
        optimizer_mem = num_params * (16 if fp16_enabled else 8)

When use fp16, total_memory = num_params * 20 ; When fp16=False, total_memory = num_params*16, this is a bug, Because the memory will be used less when using fp16 ; if you use fp16=False optimizer_mem = num_params * 12.