perf: adapt RDMA depth and recycle receive capacity

[ketor]

Change

• Select data QPs by two-dimensional (block class, depth class) geometry: scalar operations use depth 1, batches use the smallest sufficient power-of-two depth.
• Export bounded per-class opened/active/idle metrics.
• Complete replicated-MLA vLLM rank convergence by automatically enabling native CUDA-IPC same-host GET dedup; producer non-participants elide eager clients.
• Replace per-connection pull ibv_reg_mr with type-2 Memory Windows over shared chunk MRs, retaining exact-MR fallback.
• Affinitize growth chunks to their rail/NUMA node and release empty non-initial chunks after a configurable idle hold.
• Reuse one preallocated extent-zero buffer in the bounded watermark path.

No wire, slab-format, or key-schema change.

Evidence

• Local Release: 798 tests passed, 0 failed.
• Local TSan: 797 tests passed, no race report.
• xb01-0064 real RDMA: adaptive block/depth tests pass; 2,000 simultaneous depth-1 QPs complete in 20.9 s with no failures.
• Pull isolation: 202/202 connections used type-2 Memory Windows, 0 MR fallbacks on ConnectX.
• Chunk lifecycle: 3 chunks grew under 64-way 4 MiB traffic, then returned to the initial chunk; 512 MiB released.
• 64 GiB watermark: 6,000/6,000 writes, 4 bounded ticks, max tick 2.871 ms, p99 30.830 ms, max 51.304 ms.
• GLM-5.3-Flash SGLang TP8 fresh target: 813/813 KV + mamba hits; 52k-token request 205.854 -> 2.049 s.
• GLM-5.3-Flash vLLM TP4+MTP5 converged mode: fresh target 24/24 keys and 50.49 MiB/rank hit; secret verification passed; active server QPs reduced from the previous 33 to 6.

Deliberate non-changes

• Shared-CQ/fixed-worker rewrite rejected after the 2,000-QP real-verbs gate passed; no measured bottleneck justifies a high-risk Serve state-machine rewrite.
• Lock-free watermark metadata I/O rejected because the bounded lock hold is 2.871 ms and a two-phase evicting state weakens the current persistence/concurrency proof.