Expand integration & feature test coverage

[vladiant]

Summary

Adds three test groups around CvCascadeClassifier and CvHaarEvaluator::Feature::calc to lift coverage on previously partially-tested paths. All new tests follow the project's Arrange / Act / Assert convention with comments per step.

Changes

1. Multi-stage cascade integration test — test_integration.cpp

• New test: CvCascadeClassifier::train with numStages=2, maxWeakCount=3, maxDepth=2, LBP / gentle boost.
• Drives the boost outer loop for more than one stage and the recursive split path in o_cvboostree.cpp (depth > 1), plus the inter-stage sample-weight update in boost.cpp.
• Verifies stage0.xml, cascade.xml, params.xml are produced and the cascade is loadable by the public cv::CascadeClassifier.

2. HAAR feature-set integration variants — test_integration.cpp

• HAAR CORE mode produces a usable cascade — exercises the diagonal/centred Haar branch in CvHaarEvaluator::generateFeatures.
• HAAR ALL mode produces a usable cascade — exercises the 45-degree rotated branch.

3. Direct CvHaarEvaluator::Feature::calc tests — test_features.cpp

• Exposes the protected nested Feature type via a thin HaarFeatureProbe : public CvHaarEvaluator subclass.
• upright two-rect feature on a vertical-step image — explicit numeric expectation (−3200) on a known integral image.
• upright feature returns zero on a uniform image — sanity / balanced filter.
• upright three-rect feature uses rect[2] when its weight is non-zero — covers the conditional third-rectangle branch in the inline calc().
• tilted-feature branch reads the tilted integral image — drives the tilted ? _tilted : _sum branch via CV_TILTED_OFFSETS and asserts the rotated-rect area on an all-ones image (2·w·h = 32, documented in the test).
• Adds <opencv2/imgproc.hpp> for cv::integral.

Verification

[doctest] test cases: 103 | 103 passed | 0 failed | 0 skipped
[doctest] assertions: 345 | 345 passed | 0 failed |
[doctest] Status: SUCCESS!

Test-suite runtime ~26 s (was ~17 s; the extra time is the multi-stage integration test running both numStages=2 and the two additional HAAR variants).

Coverage impact

Metric	Before	After
Lines	57.5% (2687 / 4676)	57.7% (2697 / 4676)
Functions	86.4% (223 / 258)	86.4% (223 / 258)
Branches	36.8% (1840 / 5006)	37.2% (1857 / 5006)

Per-file highlights:

File	Before	After
haarfeatures.cpp	~73%	75%
boost.cpp (boost loop)	~70%	70%
o_cvboostree.cpp (depth>1)	~41%	41%

Most of the integration changes only flip a small number of branches because the existing single-stage HAAR BASIC test already covers the bulk of haarfeatures.cpp; the direct Feature::calc tests give the largest line/branch lift in this MR.

Risk

Test-only change. No production source modified. The HaarFeatureProbe subclass merely re-exposes a protected nested type for white-box testing — it does not override any virtual methods.

Checklist

• All 103 tests pass locally (clang 18 / Linux / OpenCV 4.6).
• Coverage verified with gcovr against the GCC 13 instrumented build.
• No new compiler warnings.
• AAA-with-comments doctest pattern followed throughout.