fix(harness): apply an edited max_turns, require everything=true to drop every scenario, and import fi.simulate on the callback adapter's main thread

[KarthikAvinashFI]

Part of TH-8142.

Why

• An edited max_turns was accepted by the platform editor but chat runs ignored it, because only background noise and persona fields were overlaid onto the scenario a run plays.
• A single misread builder turn could call drop_scenario('*') and replace a whole suite with what came after it.
• Every hosted chat run of a callback-style agent failed before the first turn with Plugins must be registered on the main thread: the callback adapter imported fi.simulate on a request thread, and with livekit installed that import registers plugins, which livekit allows only on the main thread. Voice agents do not use this adapter.

What changed

• max_turns is now one of the editable fields applied to the scenario a run plays.
• Dropping every scenario needs an explicit everything=true; without it the tool refuses and says so. After an explicit drop-all, the tool says the emptied suite is not saved until a scenario is submitted, and the write-scenarios skill text matches.
• The callback adapter imports fi.simulate at module level, so it runs on the main thread at process start.

Tests

• tests/harness/test_scenario_edits.py: an edited max_turns reaches the scenario.
• On a stored run whose scenario was sealed with max_turns 8 and edited to 3, the chat runner's scenario reads 3 on this branch and 8 on dev.
• tests/test_harness.py: drop-all is refused without everything=true and accepted with it.
• tests/harness/test_callback_http_adapter.py: runs the adapter as a bundle does and sends one request; fails before the fix when livekit is installed, passes after.
• tests/harness + tests/test_harness.py: the same failures as on dev (database-backed and prompt-budget tests), nothing new.