Support opensearch-neural-sparse-encoding-multilingual-v1 as OpenSearch-provided pretrained models

[zhichao-aws]

As opensearch-neural-sparse-encoding-multilingual-v1 is released at Hugging Face, we should also release it as OpenSearch-provided pretrained models to help users deploy it in OpenSearch clusters.

To release the model, we need to:

1. add a sparse tokenizer auto release pipeline.
2. enhance existing sparse model release pipeline to support model-side sparse vector prune.

[zhichao-aws]

Please feel free to assign this issue to me. Thanks.

[zhichao-aws]

To support these features, we'll need new parameters for the workflow. But the GH workflows have limits that supports at most 10 arguments. To resolve this, will replace some optional fields with a new "custom_params" field. And we'll use json string for these optional params