fix: repair TCP port-remapping slow path

[ghazariann]

Summary

Fixes #2. Four bugs in the TCP port-remapping slow path. The slow path (triggered when the virtual port is already taken) had not been tested (all existing tests ran sequentially, so the port was always free by the time the second sandbox started).

Bugs fixed

Bug 1 - _allocate_real_port held socket blocked the child's bind

The parent kept the allocated real port socket open to prevent stealing, then
rewrote the child's sockaddr to that port and sent CONTINUE. The child's bind() hit EADDRINUSE because the parent was already holding the same
address:port (only released later in PortMap.close()).
Fixed by closing it immediately, accepting the same tiny race window already mentioned in _try_reserve_port.

Bug 2 - os.pidfd_open not available in Python build

fixup_getsockname used os.pidfd_open which raises AttributeError on some builds.
Fixed by using _libc.syscall(_NR_PIDFD_OPEN) directly, consistent with the
rest of the codebase. In the future, it makes sense to separate all syscalls into
a different file like _syscall.py, since an identical _NR_PIDFD_OPEN was
found in _context.py.

Bug 3 - fixup_getsockname leaked the duplicated fd

s.detach() does not call os.close(), so local_fd was never closed on the
success path. The leaked fd kept the child's socket alive after the child
exited, holding the port bound in the parent. Fixed by moving\ os.close(local_fd) into the finally block.

Bug 4 - NotifSupervisor.stop() never called port_map.close()

Proxy listener sockets were left running after sandbox teardown. Fixed by
calling port_map.close() at the end of stop().

Tests

Two new tests added to TestPortRemap:

• test_slow_path_host_holds_virtual_port: parent process holds the virtual
  port, forcing the slow path without threads or timing dependencies.
• test_slow_path_two_concurrent_sandboxes: two sandboxes run concurrently; sandbox 1 holds the port open while sandbox 2 must remap.

[congwang-mk]

Thanks for your PR. Some comments below

[congwang-mk]

duplicated with the one in _context.py ?

[ghazariann]

Yes, I refrain from importing it from _context, should we separate all syscall wrappers into a different file, like _syscalls.py

[congwang-mk]

This could be flaky, could it be improved?

[ghazariann]

fixed in 6e68e2a