fix(simulation): prevent concurrent prepare threads

[MA1503]

Symptom

Opening Step 2 (environment setup) in the frontend while a preparation is
already running spawns another full persona-generation thread. With repeated
visits, several generators run at once — they split the Ollama throughput
between them and race on the same profile output files. The UI shows
progress counters that jump around (observed: three concurrent generators at
counter bands 30 / 139 / 159 of 252).

Cause

Two gaps in backend/app/api/simulation.py:

1. /prepare/status reports not_started while a prepare is running.
   The status endpoint only checks whether the output files already exist
   (_check_simulation_prepared). While generation is still in flight those
   files are absent, so with a simulation_id and no task_id the endpoint
   returns "status": "not_started" — telling the frontend that no
   preparation has begun.

2. /prepare has no duplicate guard. It creates a task and immediately
   starts threading.Thread(target=run_prepare) with no record of an
   in-flight prepare, so nothing stops a second (or third) call from starting
   another thread for the same simulation.

Combined, a client that polls /prepare/status, sees not_started, and calls
/prepare again keeps launching new generation threads.

Fix

A module-level registry _active_prepares maps simulation_id → task_id:

• /prepare refuses to start a second prepare while one is active for the
  same simulation and returns the existing task_id (status preparing)
  instead of spawning another thread. A stale/finished entry is cleared so a
  genuine re-prepare still works.
• The background thread is wrapped so the registry entry is removed in a
  finally block — on success and on failure.
• /prepare/status now checks the registry: when a task is PENDING or
  PROCESSING it reports preparing with the task's real progress and
  message, instead of the misleading not_started.

After a backend restart the registry is empty, so a stale preparing value in
state.json cannot block a fresh prepare.

Repro

1. Start a prepare for a simulation (Step 2).
2. While persona generation is running, open Step 2 again one or more times
   (or poll /prepare/status with only simulation_id, see not_started,
   and call /prepare again).
3. Before the fix: multiple persona-generation threads run concurrently,
   counters jump, profile files are overwritten by racing writers.
   After the fix: subsequent /prepare calls return the existing task and no
   second thread starts; /prepare/status reports preparing with real
   progress throughout.

🤖 Generated with Claude Code