[DOC] Add LAST_TOKEN pooling mode to text embedding model documentation

[aneesh-db]

What do you want to do?

• Request a change to existing documentation
• Add new documentation
• Report a technical problem with the documentation
• Other

Tell us about your request.

The pooling_mode parameter in the Register Model API documentation (_ml-commons-plugin/api/model-apis/register-model.md) currently lists supported values as mean, mean_sqrt_len, max, weightedmean, and cls.

A new lasttoken pooling mode is being added in ml-commons (opensearch-project/ml-commons#4711) to support decoder-only text embedding models (e.g., Qwen3-Embedding, GPT-style models) where the final non-padding token captures cumulative context through causal attention.

The documentation should be updated to:

1. Add lasttoken to the list of supported pooling_mode values in the model_config object table
2. Describe that lasttoken uses the last non-padding token's embedding, and is useful for decoder-only models where the final token captures cumulative context

Version: 3.4

What other resources are available?

• ml-commons issue: [FEATURE] Add LAST_TOKEN pooling mode for decoder-only text embedding models ml-commons#4709
• ml-commons PR: Add LAST_TOKEN pooling implementation for text embedding models ml-commons#4711

[kolchfa-aws]

@aneesh-db The PR is not merged yet, so is this code going into 3.6?