Implement instance pipeline

[r4victor]

Part of #3551

This PR:

• Refactors and fixes scheduled_tasks/instances.py for easier migration to pipelines.
• Implements InstancePipeline.
• Updates FleetPipeline to respect instance locks.
• Updates API and provisioning logic to respect instance locks.
• Introduces FleetModel.current_master_instance_id column for cloud clusters. Implements current master instance selection logic in FleetPipeline. Updates instance provisioning logic to use FleetModel.current_master_instance_id. (This fixes a race condition with updating siblings in case of master failure and allows re-selecting master.)

Benchmarks (SQLite, one server replica):

• Provisioning 50 AWS instances (from submission to idle) goes from ~3-4 min to ~2m, which is now about the same as provisioning one instance.
• Provisioning 200 AWS instances (from submission to idle) goes from ~11 min to ~4m, which is now mostly limited by AWS RunInstances rate limits (5 rps).
• Other instance processing was previously capped at 75 instances/min. It's now workers_num * processing_rate, e.g. typically 600-1200 instances/min for check instances depending on check network latency. With many resources, all processing including termination becomes ~10x faster.
• Pipeline throughput can further be improved by tweaking pipeline parameters, e.g. workers_num.

Notes:

• No lock_timeout errors occur on SQLite even with very high write load due to pipelines using quick sessions only.