fix(rdma): retain activated pooled endpoint

[ketor]

Problem

PR #309 activated an idle endpoint but erased it without assigning it to the caller. Every cache hit leaked the endpoint and its process-wide budget, forcing fresh QPs and eventually bounded waits.

Fix

Assign the activated endpoint before removing it from the idle pool.

Verification

• local lifecycle tests: 3/3 PASS
• local RDMA safety tests: 3 PASS, 1 hardware-dependent skip
• 0064 exact regression: broken PR fix(rdma): isolate idle endpoints before draining #309 cold C32 makespan 36.13s; fixed commit 9.33s
• 0064 fixed warm C32: 5.00s vs pre-governance baseline 5.61s
• 0064 fixed cold C32: 9.33s vs pre-governance baseline 9.64s
• fixed candidate remained healthy with no fatal/error log matches