What does this PR do?

Many changes to prepare a proper TGI image compatible with Jetstream Pytorch. In particular:

torch default type set to bfloat16 (because that's what's used mostly in TPUs anyway).
Make warmup prepare smallest size from bucket (that it was previously skipping).
ignore common warnings in pytests to avoid cluttering tests outputs
add jetstream installation step to TGI image
update TGI image (refer to commit message for details)
install torch for cpu to avoid installing the default gpu version when possible.

huggingface / optimum-tpu