fix: reduce buffer ring memory 8x + fix H2 stream leak on connection close

[FumingPower3925]

Summary

Two critical fixes for the metal benchmark regressions observed in v1.1.0-beta.9.

Fix 1: Buffer ring memory reduction (CRITICAL)

Reduce bufRingCount from 4096 to 512 per io_uring worker.

Root cause: Each worker allocates bufRingCount * BufferSize bytes of mmap'd memory for the provided buffer ring. With the Greedy preset (BufferSize=65536) and bufRingCount=4096, each worker used 256 MB. On metal instances:

Instance	vCPUs	Workers	Buffer Ring RSS	Observed RSS
c7g.metal	64	64	16 GB	17 GB
c7i.metal-24xl	96	96	24 GB	25 GB

v1.0.0 had no provided buffer ring (it used per-connection recv buffers), which is why it showed only 370-470 MB RSS.

With bufRingCount=512, each worker uses 32 MB, reducing total to 2-3 GB on metal. 512 concurrent multishot recvs per worker is more than sufficient.

Fix 2: H2 stream leak on connection close

Add Manager.Close() method and call it from CloseH2().

When an H2 connection closed, CloseH2() only cleaned up the encoder but left all Stream objects in Manager.streams map referenced indefinitely. Over many connection cycles, this leaked stream objects, headers slices, and outbound buffers.

Manager.Close() iterates the map, releases non-async streams back to the pool, and clears pending window updates.

Impact

These fixes address the reported metal benchmark issues:

• 25 GB RSS on x86 c7i.metal-24xl → expected ~3 GB
• 17 GB RSS on arm64 c7g.metal → expected ~2 GB
• Missing adaptive configs (3 configs crashed, likely OOM from combined engine memory) → should resolve
• 5-10% H1 RPS regression on ARM64 → likely caused by TLB/cache pressure from 16+ GB mmap'd regions

Test plan

• mage fullCompliance — all 4 phases pass
• All tests pass with race detector
• h2spec: 146/146 on io_uring and epoll
• Re-run metal benchmarks from benchmarks repo to verify RSS and RPS