jhm: bound the test-output echo on every reap, not just timed-out ones (#570)

[ohohoreilly]

jhm: bound the test-output echo on every reap, not just timed-out ones (#570)

jhm --test can hang forever reaping a test that has ALREADY EXITED, taking
the rest of the suite with it and burning the whole CI job with no diagnostic.
An aarch64 run sat in it for 110 minutes with 61 tests still queued, having
printed one line past the previous progress update:

TESTS: 4 running, 61 queued
TEST: orly/server/import_replication.test

No TIMEOUT:, no EXITCODE:, no echoed output.

The per-test timeout cannot catch this. Its sweep only runs when TryWaitAll()
returns nothing; the absence of a "TIMEOUT: ... still running" line proves the
test exited on its own, putting jhm in the reap where the deadline no longer
applies. It then blocked in Base::EchoOutput, which reads to EOF -- and EOF on
a pipe needs EVERY write end closed. import_replication.test forks orlyi
servers; any that outlive the test still hold its stdout.

EchoOutputBounded already exists for precisely this, written for #537, and its
comment names the failure exactly. It was simply scoped too narrowly: only the
SIGKILLed branch used it, on the assumption that only a killed test strands
children. A test that exits on its own can stand up servers and fail before
reaping them just as easily.

So use it on both paths. It costs nothing in the normal case -- a pipe whose
writers are all gone polls readable immediately and reads zero, so the quiet
period is only ever spent in the pathological case it exists for.

Verified both directions with a probe that forks a "test" which spawns a
grandchild holding stdout and then exits non-zero: the read-to-EOF shape hangs
and has to be killed, the bounded one returns in 2s and still prints the test's
output. jhm rebuilds clean via bootstrap.