Executor scheduling disabled

[calavera]

Context

We want to be able to disable scheduling in specific executors. This is useful during rollout operations where we might have executors live longer and we don't want the scheduler to schedule any work on them even if they remain alive.

What

Cordon API

Add a new internal api to cordon executors under /internal/cordon-executors.
This api can receive a list of specific executors to cordon, or cordon all present executors if the list if empty.
When the scheduler receives this request, it will propagate a new state to all executors in the request.
The request tries to wait until all executors have received the state update by spawning multiple tasks waiting for state updates in the executor.

Deficit API & warm container reallocation for cordoned executors

When an executor is cordoned, two things were broken:

1. Deficit API under-reported: count_active_idle_containers (function pools) and count_pool_containers (sandbox pools) counted containers on cordoned executors as available capacity. Since those containers can't accept new work, the deficit was too low.

2. Warm containers were stranded: The CordonExecutors handler only set executor state — it didn't mark affected pools dirty or terminate warm containers. They sat on the cordoned executor consuming resources but couldn't be used.

Changes:

• Mark affected pools dirty on cordon so the buffer reconciler re-evaluates them
• Exclude cordoned-executor containers from pool_container_count, count_active_idle_containers, and count_pool_containers
• Add Phase 0 drain in buffer reconciler to terminate idle/warm containers on cordoned executors, then let existing phases create replacements on healthy executors

Testing

• Integration tests for the cordon-executors API
• All 239 existing unit tests pass

Contribution Checklist

• I ran just fmt to format the code.
• All PR Checks are passing.