Improve CUDA driver failure diagnostics - #4903

)

Pull requests now automatically run only the fast CI jobs. The slow
Windows and pycolmap jobs run once a maintainer adds the new `full-ci`
label. On #4903, the gated jobs took about 3.4 runner-hours, and the
slowest one (the pycolmap Windows wheel, 83 min) set the time to green.

Runs automatically on every PR update:
- COLMAP (Ubuntu), COLMAP (Mac), COLMAP (Docker)
- PyCOLMAP Linux wheel (`ubuntu-24.04`) and the docs build, which
depends on it

Runs only with the `full-ci` label:
- COLMAP (Windows): `windows-2025-vs2026 Release`, `windows-2025-vs2026
Release CUDA`
- PyCOLMAP wheels: `macos-14 arm64`, `windows-2025-vs2026`,
`ubuntu-24.04 CUDA`

Details:
- The gated jobs' required status checks in the main-branch ruleset stay
"Expected" until the label is added and the jobs pass, so they still
gate merging. No ruleset changes are needed.
- The label stays on the PR, so later pushes re-run the gated jobs
automatically. Only users with triage access or higher can add labels.
- Pushes to main and release/*, releases, and manual dispatches run
everything as before.
- To gate only some pycolmap wheels, `build-pycolmap.yml` splits the
build job into `build` (Linux) and `build-gated` (macOS, Windows, CUDA).
The second job reuses the first job's name, env, and steps through YAML
anchors, so check names, artifact names, and cache keys are unchanged.
`pypi-publish` now needs both jobs.
- Adding any other label does not cancel an in-progress run, because
such label events get their own concurrency group.
- The `assigned` trigger is removed from all workflows, as it re-ran CI
whenever someone was assigned.
- `doc/contribution.rst` documents the label.

Verification performed:
- Compared the old and new workflows with YAML anchors resolved. Apart
from the intended trigger, concurrency, and `if` changes, the jobs are
identical, and the two pycolmap matrices together equal the original
matrix.
- Ran actionlint 1.7.12. It accepts the new expressions and anchors. Its
only warnings are false positives about `arch`/`deploymentTarget`
missing from the Linux-only matrix; those values were already unset for
that entry.
- This PR uses the new workflows itself: without the label, the five
gated checks should stay "Expected". Adding `full-ci` should run them.

## Summary

Merge upstream 9371f7b and PR4895
maintainer revision db7586c into
d297c1e, retaining independent stream,
blocking-sync, LoMa resize/transfer and ONNX stream changes.

Retain upstream option initialization, CUDA diagnostics and RootSIFT zero-row
safety. Use the maintainer's generic fail-closed BF16 helper and thin LoMa
wrapper. Register CUDA tests once, keep their CUDA dependency explicit, and
fix the sparse-matrix test's mixed CMake link signature.

## Validation

Incremental over d297c1e; retained the imported merge parents.

22 affected CPU objects compiled and CLI linked incrementally;
63 focused CPU tests and 7 RTX3080Ti CUDA tests pass, zero skips/failures.
Actual separate CUDA-enabled CMake configure/generation passes. The initial
ONNX test-link and Poisson CMake failures were corrected; only outstanding
test links were completed, with no repeated CLI or object build.

The one-line Poisson test CMake fix was configured after native tests and
changes no compiled production code. Public formatting changes whitespace
and include ordering only. CUDA objects are reused for declaration/comment
header deltas. Nine inherited archive identities match qualified controls;
the remaining inputs inherit the frozen native-build link closure.

No performance, memory, whole-scene quality, HIP, Python-binding or
unsupported-device end-to-end LoMa claim. PR4895 remains unmerged upstream.

AI disclosure: Totally vibe coded.

## Upstream commits

- `5ff7e73c4efa1fcad0a7766f0242ed8cec700831` Avoid lambdas as default member initializers in feature options (MSVC ODR violation) (colmap#4900)
- `3dd12895f96ae88f43eb7cfee8b26c467c6954e1` Fix CLI option names in the feature and FAQ docs (colmap#4901)
- `15526bb9a7bca6d1843b8b849169d7f6ce4f900b` Improve CUDA driver failure diagnostics (colmap#4903)
- `9371f7b47ebe660dfa8a8d968ff8def199ef0456` Preserve zero descriptors in RootSIFT normalization (colmap#4891)