Worker instance management system

Currently we rely on ec2 api and tagging to manage worker instances, for example

* Launch an instance
* Check if an instance is already running for a session
* Terminate instance for cost saving

To support more variety of agent environment, such as lambda or codeBuild, we should abstract the mechanism to avoid directly depending on ec2 apis.

## spec

- We use DynamoDB to manage the instance status per session. Running, provisioning, terminated.
- use conditional update if necessary to eliminate possible data race

This also enables to achieve #156 more easily.

Start by tracing the current EC2 API and tagging flow for launching, checking, and terminating worker instances. Define how per-session DynamoDB states and conditional updates cover running, provisioning, and terminated; done means worker management no longer depends directly on EC2 and can support other agent environments such as Lambda or CodeBuild.

Written by the indexing model from the issue text.