fix(aws): use rolling ECS deployments

[tylerslaton]

Summary

• configure ECS services to keep the existing task healthy while starting its replacement
• assert the rolling deployment percentages in the CDK test suite
• document the brief two-task overlap during deployments

Why

With a desired count of one, minimumHealthyPercent: 0 and
maximumHealthyPercent: 100 require ECS to stop the running task before it can
start the replacement. That creates an immediate Kite outage during every image
deployment.

Using minimumHealthyPercent: 100 and maximumHealthyPercent: 200 lets ECS
start and health-check the replacement before retiring the prior task. The
deployment circuit breaker and automatic rollback remain enabled.

Verification

• pnpm check-types
• pnpm test (170 tests)
• (cd agent && uv run pytest) (72 tests)
• node node_modules/railway/dist/iac/bin.js
• (cd deployment/aws && pnpm build && pnpm test) (10 tests)
• verified kite-staging, kite-prod, and kite-community remain healthy with one running task after applying the equivalent live ECS configuration