Do not split decoder checkpoint files

[dacorvo]

What does this PR do?

Neuron 2.18 introduced support for loading weights in SafeTensors format. Due to the nature of SafeTensors, it is no longer essential to split model weights to conserve memory. Weights can be loaded directly from .safetensors checkpoints and then loaded to Neuron cores directly.

Implements #565

Edit:

• I modified a bit one of the tests to use a cached model and save compilation time,
• I had to fix the docker build broken by a recent release of setuptools

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.