VRAM detection now uses cf-orch free VRAM when coordinator is running,
making the scheduler cooperative with other cf-orch consumers.
Enqueue return value now checked — queue-full tasks are marked failed.
Replaces the spawn-per-task model for LLM task types with scheduler
routing: cover_letter, company_research, and wizard_generate are now
enqueued via the TaskScheduler singleton for VRAM-aware batching.
Non-LLM tasks (discovery, email_sync, etc.) continue to spawn daemon
threads directly. Adds autouse clean_scheduler fixture to
test_task_runner.py to prevent singleton cross-test contamination.
