comp_feat: fix flaky parallel exit-255 (atomic resolved-config write)

[lenzo-ka]

Root cause (confirmed)

train-g2p-lda-vtln and train-parallel intermittently failed at module 000
comp_feat with ERROR: Parallel job failed with exit code 255 and no cause in the
step log. The failing shard's error was written to a per-shard qmanager/*.err file
that CI did not archive; capturing it (commit 1) revealed the real fault:

malformed JSON string, neither array, object, number, string or atom, at character
offset 0 (before "(end of string)") at .../SphinxTrain/Resolved.pm line 199.
BEGIN failed--compilation aborted at .../make_feats.pl line 51.

Every script that loads SphinxTrain::Config calls Resolved::sync_runtime() at
BEGIN, which regenerates etc/sphinx_train.resolved.json when it looks stale.
write_file() opened the target with >, truncating it in place and bumping its
mtime. When the parallel comp_feat shards start together, one shard truncates the
file (making it look fresh to a sibling via the new mtime) while a sibling reads it
in that window — reads an empty file — and dies parsing empty JSON at BEGIN,
before writing any log. The fail-fast then aborts the stage as exit 255. Fast
machines win the race (it never reproduced on a 36-core box, nor at 2 CPUs / 12-way
oversubscription); the slower GitHub runners lose it, which is why it looked flaky and
runner-specific rather than a code bug. comp_feat is module 000, so this has nothing
to do with G2P/phonetisaurus, and nothing to do with any recent merge.

Commits

1. ci: archive qmanager on the AN4 jobs — captures the per-shard errfile that
   was the only record of the real error. This is what pinned the root cause.
2. WaitForScript: don't turn an already-reaped shard into a spurious failure —
   waitpid() returning -1 made $? >> 8 == -1, an exit 255 with no failed
   job. Robustness, independent of the root cause.
3. comp_feat: create output dirs under the warp folder, race-tolerantly —
   make_feats.pl created warp dirs under the base feature dir and used a bare
   mkpath() that dies File exists under a parallel sibling. Correctness +
   robustness, independent of the root cause.
4. Resolved: write the resolved config atomically (the fix) — write a
   per-process temp file and rename() it over the target (atomic on POSIX), so a
   concurrent reader always sees a complete file, never a truncated one.

Validation

All scripts pass perl -c. A full AN4 G2P+LDA+VTLN train (VTLN 0.90–1.10, NPART=2,
Queue::POSIX) with the patched runtime completes comp_feat with all 14 shard logs and
all five warp dirs fully populated (948 mfcs each) and trains through normally — no
regression, and warp features now land in the correct feat/<warp>/ directories.
Because the flake is timing-dependent on the GitHub runners, this PR's own CI run is
the real test: with the atomic write, train-g2p-lda-vtln and train-parallel
should be green.