TGI improvements

[dacorvo]

What does this PR do?

This brings several fixes and improvements to NeuronX TGI.

Improvements

• Use model default generation parameters like top_k and top_p when they are omitted by the client,
• avoid loading model twice when exporting (divides deployment time by almost two),
• make sure MAX_BATCH_SIZE is set on SageMaker.

Fixes

• TGI NeuronX DLC (Optimum-neuron) 0.0.20: SageMaker deployment failure with llama-2 7B #516 : warmup phase did not end cleanly with some configurations.

[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.