Added support for launching inference endpoint with different model dtypes

[shaltielshmid]

Following the documentation here: https://huggingface.co/docs/text-generation-inference/en/basic_tutorials/launcher#quantize

If the dtype specified doesn't match any of our conditions, we don't specify anything and let it fallback to the default.

[clefourrier]

I need to test it in depth but from a first glance, it LGTM, it will work well with our logging system too, which is nice. Good job!
For next time, it would be great if you could check is someone else is already working on the related issue, to avoid duplicate work 🤗

Will fix #117 , linked to #120 .

[clefourrier]

LGTM!