reactor: Teardown never disposes live or pooled TcpConnections, so every Stop() leaks their native write slabs

[MDA2AV]

Summary

Reactor.Teardown frees the ring, the provided-buffer slabs, the UDP blocks and the timespecs, and CloseAcceptedTcpSockets closes the accepted descriptors - but neither _pool nor the TcpConnection objects still in _connections are disposed. Each one owns a NativeMemory.AlignedAlloc'd write slab (WriteSlabSize, 16 KiB by default; more when grown, plus the iovec/msghdr in Segmented mode and the per-connection buffer ring in incremental mode), and TcpConnection has no finalizer, so that memory is unreachable after Stop().

Up to PoolMax x WriteSlabSize per reactor per stop - 16 MiB at the defaults. It matters for hosts that start and stop many servers in one process, which is exactly what the test suites do.

Where

• src/ioxide/Reactor/Reactor.Runner.cs Teardown()
• src/ioxide/Reactor/Transport/Tcp/Reactor.Tcp.cs CloseAcceptedTcpSockets() - closes the fd, drops the object
• src/ioxide/Connection/Tcp/TcpConnection.cs Dispose() - only reached from Recycle when the pool is already full

Fix

In Teardown, after the ring fd is closed (so no SEND can still reference a slab): Dispose() every non-null _connections[fd] and drain _pool through Dispose(). DisposeIncremental goes with them.