fix(token-counter): count literal special-token markers as text

[ducanhnguyen223]

Summary

• use encode_ordinary() in TiktokenCounter so literal tiktoken markers in user text and tool schemas are counted as ordinary text
• add regressions for both message and tool-schema paths
• add a concise release note

Verification

• upstream behavior reproduced: Encoding.encode("literal <|endoftext|> marker") raises ValueError; encode_ordinary() returns 9 tokens
• targeted token-counter suite: 14 passed
• Ruff and git diff --check pass

The repository Hatch command is not installed in this environment, so the targeted checks were run with uv; no full suite or type check is claimed.

[vercel]

Someone is attempting to deploy a commit to the deepset Team on Vercel.

A member of the Team first needs to authorize it.

[CLAassistant]

All committers have signed the CLA.