fix(inventory): B2: take a lease without waiting on the group's other writers

[dannash100]

take_lease ran 1s to 60s+ in prod while the read endpoints stayed at 0.3s, past the plugin's 15s timeout, so runs fell back to a cache holding no secret values.

• InventoryLease::take locked the group row FOR UPDATE, which conflicts with the FOR KEY SHARE Postgres holds there for every child-row write; ~25 tables reference server_groups.
• FOR NO KEY UPDATE instead: FK writes stop blocking it, two concurrent takes still queue.
• Mechanism and fix are proven; that this is what prod waited on is not, and prod hasn't been re-measured.

🦸 Review Hero

• Run Review Hero