fix(rdma): v2.6.1 hardening — revert unilateral, fix MR leak, drop per-request logs

[ketor]

Release-blocking fixes (review of v2.6.0)

1. Revert unilateral PUT (RoundTrip) to dual-sided WRITE_WITH_IMM

Unilateral never entered the standard client path: KVClient::Put/BatchPut both route through CacheFrom → CacheMany (pipelined), so the advertised performance path did not match the actual call chain. Removes PostSendNotify + nbuf_/nmr_ entirely.

2. Fix per-QP notification MR/memory leak

Close() cleared nmr_/nbuf_ vectors BEFORE the dereg/delete loops, so both cleanup loops were always empty → every QP teardown leaked 64B/slot buffers + MRs. nbuf_/nmr_ removal eliminates the leak; remaining vectors now clear after their loops.

3. Drop per-request INFO logging + duplicate decoders + type-punned load

• Every RDMA completion logged at INFO with global-mutex fprintf (546)
• Three copies of the unilateral decoder existed in decode_request; two removed, the third removed now
• reinterpret_cast<const uint64_t*> raw wire load removed

4. Remove dead reaper

StartReaper/StopReaper had zero callers, reaper_enabled_ unused, Close() never called StopReaper() — latent lifetime hazard.

5. Default idle back to 10 min

DFKV_RDMA_IDLE_MS default 600000; K8s launcher exports 30000 explicitly.

Verification (0064, B200, mlx5_0; 32GB recv segment, 32GB cap, 64KB slab granularity, credits=256)

Matrix	fails
t32 PUT+GET 10k, batch 1/2/8/32	0
depth 1/4/8 (t32, b1, 10k)	0
1MB PUT+GET (t32, b1, 2k)	0
evictions	0

Release-claimed t32/b1 ~83% fails root cause: double-PostRecv RQ misalignment (fixed in #266) + test-env slab eviction at 1MB granularity — NOT shared-QP WR interleaving (Acquire is mutex-guarded; no thread ever holds the same Conn*).