hostinet: drain recv buffer before EOF

[russellromney]

Summary

• remove the early RClosed() EOF return in hostinet RecvMsg
• let the socket buffer read path return buffered bytes before EOF
• add a TCP regression client for data-plus-FIN close timing

Fixes #1367.

Root cause

When a peer sends data and FIN together, qvisor can process the file-read data completion and then the EOF completion before the guest application enters recv(). The receive buffer already contains bytes, but RecvMsg checked buf.RClosed() before calling ReadFromBuf, so it returned clean EOF without consulting the buffer. Retrying on the same fd repeated the same early EOF shortcut.

SocketBuffIntern::Readv already has the correct ordering: copy buffered data first, then report EOF only when the buffer is empty and the read side is closed. This patch routes closed sockets through that path.

Verification

On the DigitalOcean KVM repro host:

runc    close=immediate ok=16 empty=0 other=0
runc    close=delay50ms ok=16 empty=0 other=0
runsc   close=immediate ok=16 empty=0 other=0
runsc   close=delay50ms ok=16 empty=0 other=0
quark   close=immediate ok=16 empty=0 other=0
quark   close=delay50ms ok=16 empty=0 other=0

Additional fixed-Quark immediate-close image checks with the same static repro client:

debian:bookworm-slim ok=16 empty=0 other=0
ubuntu:24.04         ok=16 empty=0 other=0

Regression test binary under fixed Quark:

tcp_fin_recv: 32 iterations ok

Build note: the repro host needed the existing local qkernel compile shims for drivers::tee / sync_unsafe_cell, unrelated to this networking change.