fix(rdma): isolate idle endpoints before draining

[ketor]

Change

• add explicit ACTIVE/IDLE/RETIRE_REQUESTED/DRAINING lifecycle
• pooled Acquire atomically claims IDLE before CQ use
• Release publishes IDLE only after WR completion and MR generation refresh
• SIEVE eviction atomically isolates IDLE before removing it from cache
• keepalive claims IDLE as ACTIVE before posting
• every failure/destructor path enters DRAINING before endpoint destruction and budget return

Invariant

No endpoint accepts a new operation or keepalive after RETIRE_REQUESTED. Resource credit returns only after RcEndpoint destruction.

Verification

• lifecycle activation-vs-retire race test
• retired endpoint cannot reactivate
• active failure drains directly
• RdmaSafety filtered tests

No environment is changed by this PR.