Add pytorch/inference/{cpu,gpu}/2.3.1/transformers/4.48.0/py311

[alvarobartt]

Description

This PR bumps the version of the huggingface-inference-toolkit to 0.5.4 to release the latest PyTorch DLC for Inference, that comes with bumped versions for transformers, diffusers, huggingface_hub and accelerate.

The dependency bump for transformers mainly introduces new architectures such as ModernBERT, ColPali, Falcon 3, etc., as well as several fixes and improvements overall. Also diffusers comes with new text-to-image pipelines for SANA and Flux Control.

Read more about the latest releases for each dependency on their respective release notes:

• transformers at https://github.com/huggingface/transformers/releases/tag/v4.48.0
• diffusers at https://github.com/huggingface/diffusers/releases/tag/v0.32.2
• accelerate at https://github.com/huggingface/accelerate/releases/tag/v1.2.1
• huggingface_hub at https://github.com/huggingface/huggingface_hub/releases/tag/v0.27.0

---

Additionally, this PR updates the entrypoint.sh for it to be more robust and consistent in formatting, while also adding the requirements.txt installation when the HF_MODEL_DIR is set, as when running the PyTorch Inference DLC on GKE using a mount or a volume with custom code, the requirements.txt where not being installed and custom code with custom requirements couldn't be used, but now the requirements.txt will be installed if a path in HF_MODEL_ID, HF_MODEL_DIR or AIP_STORAGE_URI is provided.

Finally, it also adds flash-attn as a dependency for the GPU image, so as to benefit from it, which is useful in some scenarios as e.g. answerdotai/ModernBERT-base as described in https://huggingface.co/answerdotai/ModernBERT-base#usage; among may others.