Avoid ONNX session load for token counts

[Mirochill]

All Submissions:

• Have you followed the guidelines in our Contributing document?
• Have you checked to ensure there aren't other open Pull Requests for the same update/change?

New Feature Submissions:

• Does your submission pass the existing tests?
• Have you added tests for your feature?
• Have you installed pre-commit with pip3 install pre-commit and set up hooks with pre-commit install?

Summary

Fixes #584.

This updates lazy token_count() paths so they load tokenizer files from the resolved model directory without initializing ONNX sessions. The change covers text embeddings, cross-encoder rerankers, sparse BM42, late-interaction embeddings, and multimodal late-interaction wrappers.

The lazy-load regression tests now exercise token_count() before embedding/reranking and still assert that the ONNX model session has not been loaded yet.

Validation

• git diff --check
• git diff --cached --check
• git show --check --stat --oneline HEAD
• Static search of the targeted token_count() paths for remaining ONNX-session loading calls
• Not run locally: tests, docs build, linter, pre-commit