docs: add HuggingFace token classification offset_mapping documentation

## Summary

KServe v0.18 added `offset_mapping` output support for HuggingFace token classification tasks. The existing token classification doc (`docs/model-serving/predictive-inference/frameworks/huggingface/token-classification/token-classification.md`) does not mention this feature.

## What needs to be documented

- How to enable `offset_mapping` output for token classification tasks
- What `offset_mapping` provides (character-level position information for each classified token)
- Use cases: NER and span-extraction applications
- Example request/response showing offset_mapping output

## Related PR

- https://github.com/kserve/kserve/pull/5244

## Since

KServe v0.18

Start by reading docs/model-serving/predictive-inference/frameworks/huggingface/token-classification/token-classification.md and the related PR 5244. Document how to enable offset_mapping, explain its character-level token positions and NER/span-extraction uses, and add a request/response example showing the output.

Written by the indexing model from the issue text.