fix(core): keep tool call pairs during token truncation

[林SO (Linxiushen)]

Why are these changes needed?

TokenLimitedChatCompletionContext removes messages from the middle of the history one at a time. When the midpoint lands on either side of an assistant tool call and its execution result, this can leave an invalid history containing an orphaned result or call.

This change treats adjacent assistant tool-call and execution-result messages as an indivisible pair during token truncation. It applies the same behavior to both explicit token_limit counting and model-client remaining_tokens counting.

Regression coverage exercises both possible midpoint positions across both token-limit modes.

Related issue number

Closes #7955

Checks

• I've included any doc changes needed for https://microsoft.github.io/autogen/. No documentation changes are needed for this internal bug fix.
• I've added tests (if relevant) corresponding to the changes introduced in this PR.
• I've made sure all auto checks have passed.

Local checks:

• pytest tests -q (230 passed)
• ruff format --check src tests
• ruff check src tests
• mypy --config-file ../../pyproject.toml --exclude src/autogen_core/application/protos --exclude tests/protos src tests
• pyright --pythonpath ../../.venv/Scripts/python.exe

[林SO (Linxiushen)]

@microsoft-github-policy-service agree

[Anuj Bolewar (anujbolewar)]

Keeping an adjacent tool-call and its result together during truncation is the right call, since splitting the pair leaves a dangling result the model cannot interpret. The helper chooses a middle message and expands in either direction, which is reasonable. One edge: when the middle message is a tool call with no following result, or a result with no preceding call, it falls back to popping a single message — consider a comment spelling out that a lone message may legitimately be removed in those cases so it reads as intended rather than accidental.