[fix] use inference pool client instead of starting one for validation

[Jackmin801]

Note

Low Risk
Small wiring change limited to validation client selection; primary risk is validation hitting a different/updated backend than before, but no auth/data-path logic changes.

Overview
Validation rollouts now use inference_pool.clients directly instead of a cached clients variable, ensuring validation always targets the current inference pool (including newly discovered servers in elastic mode).

This also removes redundant local reassignments of clients after inference pool setup/ready checks.

^{Written by Cursor Bugbot for commit b6a7b28. This will update automatically on new commits. Configure here.}