Edit benchmark code

[wangd12rpi]

• If model name contains mamba, use MambaForCausalLM. https://huggingface.co/state-spaces/mamba-2.8b-hf
• Return seconds per question
• return GPU memory usage, use torch.cuda.max_memory_allocated() // 1024 // 1024 // 1024

[CalissaT]

still need to add mamba check and modify other files. only integrated in fiqa right now 11/19