libkernel: writing to a pipe whose reader is closed returns EINVAL on Windows instead of EPIPE; fix branch ready #5246

On Windows, write, writev, sceKernelWrite and sceKernelWritev to a pipe whose read end has been closed fail with EINVAL instead of EPIPE. The host C runtime maps ERROR_NO_DATA (232) to EINVAL, and FreeBSD 11.4 pipe_write returns EPIPE. A guest that treats EPIPE as "the reader went away" (the usual way a writer ends cleanly) sees a generic invalid-argument error instead.

Fix branch: https://github.com/husseinkad/AnyPS5/tree/fix/libkernel-pipe-write-epipe (one commit on a recent main: File/src/Open.cpp NativeWrite, the TransferIovecs error path in File/src/Stdio.cpp, and the new test tests/GuestPipeClosedWrite.cpp, registered as guest_pipe_closed_write). ERROR_NO_DATA is now remapped to EPIPE in both places. NativeTransfer is left alone because #4885 changes it.

Checks:

• The new test fails on main (abort) and passes with the change; guest_file_vector, guest_kernel_fcntl and guest_dup pass.
• Probed through the guest exports against main: a write of 1, 4096, 4097 bytes and 1 MiB, and writev with 1 to 4 iovecs (including zero-length ones), now give EPIPE (32) for blocking and nonblocking pipes, and sceKernelWrite/sceKernelWritev return 0x80020020. A 1 MiB write blocked on a 4096-byte pipe wakes with EPIPE when the reader closes (with two blocked writers main returned EPIPE for one and EINVAL for the other; now both get EPIPE), with no hangs. Unchanged against main: a full pipe (EAGAIN or blocking), a write to the read end and to a closed or invalid fd (EBADF), and writes to a regular file, a read-only file and a socket. A direct probe through both ucrtbase.dll and msvcrt.dll shows the same EINVAL with _doserrno 232, which matters because local toolchains and CI use different C runtimes.
• Two independent reviews found nothing wrong with the change. check_conventions.py is clean and pr_overlap.py finds no overlap; fix(libkernel): return EAGAIN for nonblocking vector pipe I/O on Windows #4885 edits NativeTransfer nearby, so a small rebase of the Stdio.cpp hunk may be needed.

Not covered: no SIGPIPE is raised on Windows (already recorded in docs/dev/debt/libkernel.md); a write that is part done when the reader closes returns EPIPE instead of the partial count through the single-buffer path (the iovec path does return the count); the check reads _doserrno without clearing it first. A zero-length write and a zero-length writev on a broken pipe return 0 where FreeBSD returns EPIPE, and a nonblocking writev on a full pipe gives ENOSPC rather than EAGAIN; both are the same on main. Only Windows was built and run.

AI-assisted: yes