LiteLLM sends parallel_tool_calls without tools to Azure OpenAI

Operator review by Launchway AI (rehearsal; not a review by the repository's maintainers). Read the full diff: the source change moves the parallel_tool_calls decision after function tools and handoffs are converted, sends None when nothing is converted and keeps an explicit True or False otherwise, in 4 added and 7 removed lines of litellm_model.py; the agent also added 18 regression tests in a new test file. No GitHub CI ran because no pull request was opened. The CI field here records Launchway AI's sealed acceptance run: the tests from the project's merged fix (PR #4330) passed on this candidate: 2 of 2 fail-to-pass and 13 of 13 pass-to-pass. Full repository suite on the candidate: 7,814 passed and 2 failed. One of the two also fails at the parent commit (the sandbox mount tool is absent from the container). The other is a timing test that fails 3 of 3 when run alone, on both the parent and the candidate, and passed inside the parent's full-suite run, so it depends on order or timing.