bug: `fcntl(F_DUPFD)` reports `EBADF` instead of `EMFILE`

[Shounak-Ghosh]

Component: src/rawposix/src/fs_calls.rs (fcntl_syscall, F_DUPFD and F_DUPFD_CLOEXEC arms)
Severity: POSIX conformance
Found by: CONC-005a (#1304), conc_005_fd_exhaustion_isolation.c

Symptom

A cage at its virtual fd table limit cannot be distinguished from one that was handed
an invalid descriptor: both return EBADF. POSIX requires EMFILE for the former.
Well-behaved applications that retry after closing something instead treat the fd as
bad and give up.

Root cause

// src/rawposix/src/fs_calls.rs, F_DUPFD arm (~1451) and F_DUPFD_CLOEXEC arm (~1470)
Err(_) => return syscall_error(Errno::EBADF, "fcntl", "Bad File Descriptor"),

vfd_arg has already been validated by _fcntl_helper before this point, so the only
way the allocation can fail is a full table, which is EMFILE by definition.

Proposed fix

-                Err(_) => return syscall_error(Errno::EBADF, "fcntl", "Bad File Descriptor"),
+                // `vfd_arg` was already validated by `_fcntl_helper` above, so the only
+                // way this allocation fails is a full virtual fd table: that is EMFILE,
+                // not EBADF. Reporting EBADF here made an exhausted cage look like it
+                // had been handed a bad descriptor.
+                Err(_) => return syscall_error(Errno::EMFILE, "fcntl", "Too many open files"),

Apply to both the F_DUPFD and F_DUPFD_CLOEXEC arms.

Guest-visible change: fcntl(fd, F_DUPFD) on a full table returns EMFILE, was
EBADF. This is a correction toward Linux.

Un-skip on merge

Remove conc_005_fd_exhaustion_isolation.c from skip_test_cases.txt.