Add containers/tei/{cpu,gpu}/1.5.1

[alvarobartt]

Description

This PR adds a new container for TEI v1.5.1 recently released (see the release notes at https://github.com/huggingface/text-embeddings-inference/releases/tag/v1.5.1).

The main features within TEI v1.5.1 w.r.t. TEI v1.4.0 are the following:

• Added /similarity endpoint for sentence-similarity tagged models in the Hugging Face Hub at https://huggingface.co/models?pipeline_tag=sentence-similarity
• Added ONNX runtime for CPU deployments, improving the throughput
• Set CLS as the default pooling method for BERT variants (it was impacting the deployment success rate on Google Cloud as per Low deployment success rate using the HF text-embedding-inference container, with missing pooling and tokenizer #115

Note

Note that this PR also includes the changes from the https://github.com/huggingface/text-embeddings-inference/releases/tag/v1.5.0 release.

To inspect the changes required to make the TEI container work in GCP, see the diff at:

• TEI on GPU: b1d104e
• TEI on CPU: 488d642

[alvarobartt]

Note

This PR is on hold, since the CPU version requires the model to have ONNX compatible weights, and there are a bunch of models that only contain the safetensors weights.