docling-rs: add build-docling-rs.yml for riscv64 wheels - #4276

cibuildwheel evaluates CIBW_TEST_ENVIRONMENT when it creates the test
virtualenv, before CIBW_TEST_REQUIRES is installed, so importing
onnxruntime there failed with ModuleNotFoundError. Compute the path in
CIBW_TEST_COMMAND instead, once onnxruntime is installed.

In run 42497562052, the manylinux build job failed after ~1.5h with
"Failed: DID NOT RAISE KeyboardInterrupt" in
test_ctrl_c_interrupts_native_chunking (234 passed, this was the only
failure). The test fires one SIGINT 0.5s after starting to chunk a
3000-paragraph document and expects KeyboardInterrupt within 5s, but
PyChunkStream::__next__ only calls check_signals() when its 100ms poll of
the chunk channel times out with nothing queued, and the whole chunk
(~60s on this host, per the log timestamps) ran without that gap ever
appearing, so the one-shot signal was never observed.

Patch the test to re-issue the SIGINT every 0.2s until the call is
interrupted, matching the same disposition already used for kuzu's
test_connection_interrupt (patches/kuzu/0.11.3/0002-*), and drop the fixed
5s completion budget that assumed a one-shot signal lands immediately.
Wire the patch into build-docling-rs.yml: CIBW_TEST_SOURCES copies the
tests from the top-level checkout, not the sdist unpacked for the build,
so the patch applies to that checkout.

_publish-wheel.yml now requires a matrix input (main's "publish:
Validate effective wheel targets"), but this PR branch forked before
that landed and docling-rs has never been merged to main, so
build-docling-rs.yml never picked up the matching one-line update every
already-published package's workflow got. The merge ref combines the
new required-input _publish-wheel.yml with this branch's old
build-docling-rs.yml, so any run has failed startup with "Input matrix is
required, but not provided".

Pass it the same way the mass-update commit did for every other
package's publish job.

…n leak

Run 42731217360's manylinux build job failed with cibuildwheel exit code
130 (SIGINT) during pytest's own session teardown (gc.collect() in
_ensure_unconfigure), right after test_ctrl_c_interrupts_native_chunking
-- the last item in the file -- leaving it uncounted in "16 passed, 4
skipped" instead of passing.

The 2026-11-09 fix for this test's single missed SIGINT re-issues it
every 0.2s from a background thread until the call is interrupted, but
that thread checks `stop` and calls raise_signal() as two separate,
unsynchronized steps. It can commit to one more raise_signal() in the
instant before the main thread observes `stop` and returns from
timer.join(); that extra SIGINT is still pending delivery to the main
thread at that point and, with no later test in the file to catch it,
surfaced as an unhandled KeyboardInterrupt during session teardown
instead.

time.sleep() is itself interruptible, so give the pending signal a
bounded 0.2s window to land and absorb it right there in the test's own
finally block, instead of leaving it for whatever the session runs next.