Add round-trip serialization tests for *Params types

[vladiant]

Summary

The cascade trainer persists every stage's configuration to params.xml
via cv::FileStorage, then reads it back when resuming training or
emitting the final cascade. This MR locks down that contract with a new
test file that round-trips each *Params type through an in-memory
FileStorage and verifies the result matches the source — plus a focused
set of negative cases for the documented validation paths.

Library line coverage rises from 41.1% → 42.4%, with notable gains
in features.cpp (85% → 96%), haarfeatures.cpp (67% → 73%) and
boost.cpp (63% → 70%). No production code changes.

Changes

traincascade/test/test_serialization.cpp (new)

19 doctest test cases, all strictly following the Arrange / Act /
Assert pattern with comments per step. Two small helpers wrap each
instance under a named section since the write/read methods emit
children directly without their own enclosing node:

template <typename Params>
std::string writeToMemory(const Params& src, const std::string& section);

template <typename Params>
bool readFromMemory(Params& dst, const std::string& xml,
                    const std::string& section);

Both use FileStorage::MEMORY — no temp files, no I/O.

Type	Tests	Coverage
CvCascadeParams	6	HAAR / LBP / HOG round-trips; empty node; zero width; unknown feature type string
CvFeatureParams	2	Round-trip; empty node returns false
CvHaarFeatureParams	5	BASIC / CORE / ALL round-trips; unknown mode; missing mode
CvLBPFeatureParams	1	Default round-trip (256 / 1)
CvCascadeBoostParams	5	GENTLE / DISCRETE / REAL+LOGIT round-trips; out-of-range minHitRate throws; unknown boost type throws

CMakeLists.txt

• Add test/test_serialization.cpp to the test_traincascade target.

Library bug surfaced (not fixed in this MR)

CvCascadeBoostParams::CvCascadeBoostParams(int _boostType, ...) writes
boost_type = cv::ml::Boost::GENTLE after delegating to the base
constructor, so the _boostType argument is effectively ignored. The
DISCRETE and REAL/LOGIT round-trip tests work around this by
setting boost_type directly after construction; the workaround is
documented in test comments. Recommended follow-up: drop the override or
honor the argument.

Test results

[doctest] test cases:  80 |  80 passed | 0 failed | 0 skipped
[doctest] assertions: 272 | 272 passed | 0 failed |
[doctest] Status: SUCCESS!

ctest: 100% tests passed, 0 tests failed out of 1

Coverage impact (lib)

Metric	Before	After
Lines	41.1% (1921 / 4676)	42.4% (1984 / 4676)
Functions	76.0% (196 / 258)	77.5% (200 / 258)
Branches	25.6% (1280 / 5006)	27.7% (1388 / 5006)

Per-file gains:

File	Before → After
features.cpp	85% → 96%
haarfeatures.cpp	67% → 73%
boost.cpp	63% → 70%

The new tests fully exercise CvFeatureParams::read/write (including
the empty-node guard), all three CvHaarFeatureParams modes plus their
two failure branches, and both CV_Error paths inside
CvCascadeBoostParams::read (unknown boost type and out-of-range
values).

Risks / Notes

• All tests use in-memory FileStorage — no filesystem state, no test
  fixtures to clean up.
• Runtime contribution is well under one second; CI timing is
  unaffected.
• No changes to library code in this MR; the boost_type issue noted
  above is documented in test comments and tracked separately.

Checklist

• All 80 tests pass locally (ctest --output-on-failure)
• AAA pattern with comments in every test
• Happy paths and edge/error cases covered
• No new resource files committed
• No changes to library code