Bound the pty hangup reads with an alarm

[alanhc]

Closes #275.

The hangup checks in tests/test-pty.c read from a master whose slaves have
all closed. Linux answers EIO there, but on a tree without the support nothing
fails the read at all -- elfuse's keepalive slave holds the host pty open, so
the host has no reason to. The read parks, and a run against such a tree stalls
instead of reporting.

tests/test-devpts.c already takes the position that a test guarding a fix has
to fail when the fix is absent, and the readv case at the bottom of this file
does too: signal(SIGALRM, ...) plus alarm(10) around the blocking call,
cleared with alarm(0) once it returns. The first hangup block never got the
same treatment. This applies it there.

Both reads sit in one window. The second is the wedge; the drain above it reads
the same hung-up master with no deadline of its own. Hoisting the second read
out of EXPECT_TRUE is what lets the alarm be cleared before the verdict is
printed, as the readv case does.

Two smaller things came out of measuring it. The handler was silent, and its
_exit(2) discarded whatever stdout still held, so a wedged run through a pipe
produced no output at all -- neither the results collected so far nor a word
about why it stopped. It now names the timeout on stderr, and stdout is
line-buffered, so the log ends with every check that did complete followed by
the reason the run went no further.

Evidence

Measured against a tree with proc_pty_master_hung_up() stubbed to return false, which is what "without the support" means here.

	before	after
outcome	still running at a 60s cap, killed	exit 2 at 10s
output	nothing	37 lines of results, master reports POLLHUP once the slave closes FAIL, then the timeout line

After, on the stubbed tree:

  packet-mode read carries a status byte OK
  master reports POLLHUP once the slave closes FAIL: no POLLHUP after the last slave closed (errno=2)
  queued output survives the hangup OK

test-pty: TIMEOUT waiting on a hung-up master read

On an unmodified tree the file still reports 70 passed, 0 failed - PASS,
exit 0, in 3 seconds.

make check-format passes (clang-format 22.1.8, commentflow), make indent
leaves the file unchanged, and the cross build is clean under -Wall -Wextra.
A full make check did not run locally for want of aarch64-none-elf-as, so
everything outside test-pty is on CI.

---

Summary by cubic

Fixes #275 by bounding the hung-up master reads in tests/test-pty.c with an alarm: on trees without hangup support those reads previously blocked forever, now the run exits 2 at 10 seconds with a timeout message.

Bug Fixes

• Applies the existing SIGALRM + alarm(10) pattern from the readv case, clearing the alarm before printing the verdict.
• Writes the timeout reason to stdout, not stderr (test-matrix.sh discards stderr), and line-buffers stdout so the log retains all completed checks when the alarm fires.

^{Written for commit 2848eb0. Summary will update on new commits.}

[jserv]

One note outside the diff: the description says "It now names the timeout on stderr", but the handler writes STDOUT_FILENO and the new comment argues at length for stdout. One word to fix in the prose.

[jserv]

The alarm stays armed while the drain verdict prints. If the drain read burns most of the window, SIGALRM can fire after that read already returned, and it can land between TEST()'s label (printed without a newline, so still buffered) and the newline EXPECT_TRUE adds. The handler comment claims both guarded reads sit at a line boundary; that holds for the reads, not for the gap between them, and the TIMEOUT line would then name a read that did not block. Clearing at line 850 and re-arming just before the EIO read costs two lines.