Understand why jobs are still round to drained fleets

[huydhn]

As part of pytorch/pytorch#190347 deployment, we drained ue2 https://github.com/pytorch/ci-infra/actions/runs/29603818377/job/87962059939. Previously, draining a fleet correctly blocks jobs going there. However, the behaviour seems to have been changed in which jobs are still being routed to the drained cluster as shown in the dashboard

Per @jeanschmidt, LF fleet draining still works correctly.

Q:

1. Could something in OSDC cause this? Or is this something that GitHub changes recently?
2. Shoud we implement runner group routing decision ourselves on CI? Similar to mt <-> lf routing