Update modeling_mpt.py

[madhavatreplit]

Why

We had an open source PR on the HF repo that enables 8-bit and 4-bit quantization: https://huggingface.co/replit/replit-code-v1-3b/discussions/19/files

What changed

Replicated those PR's changes here.

Testing

Tested that loading in 8-bit and 4-bit quantization work.

Rollout

• [x] This is fully backward and forward compatible