Include tools changes in existing turn analytics - #50943

Turn analytics currently cannot show how often the available tools change during a session. Include tools_change_count on codex_turn_event, next to the existing sampling and retry counts, so the backend can break it down using the event's usual originator/client tags.

Compare the existing shared Prompt.tools list at the common inference entry point. Increment once whenever it differs from the previous inference call, including across turns, and report the count on the current turn. The first call, identical retries and compaction do not increment it. This counts all tool-definition differences, including fields that are not serialized for some models; it measures changes to tools and does not establish an actual cache miss.

Validation: just fmt and git diff --check passed. The existing core turn-profile test covers baseline, unchanged retries, tool fields, additions, ordering, removal and continuity after a transport fallback. Analytics serialization was tested in the preceding PR revision.

just test -p codex-core: 3,694 passed, 22 failed, 9 skipped. The tools-change regression passed; the 22 failures are exactly the same tests as the saved run before this revision (no new failures).