[SAMPLE_CONSENSUS] Refactor Ellipse3D optimization kernel to internal namespace

[AlrIsmail]

Summary
Refactor SampleConsensusModelEllipse3D to move the heavy Eigen::LevenbergMarquardt optimization
kernel into a non-templated source file. See #6428 for more.

• Build Analysis: clang_build_analyzer_baseline.txt vs clang_build_analyzer_after.txt ~24% reduction
• Runtime Benchmarks: Measured over 1000 iterations on a 200-point cloud, optimization time dropped from 2353 µs to 1913 µs (~18.7% faster). Unit tests also improved from 1.66 s to 1.28 s.

Technical Changes

• Encapsulated LevenbergMarquardt in pcl::internal::optimizeModelCoefficientsEllipse3D.
• Math helpers moved to pcl::internal as inline header functions for performance.
• Added explicit template instantiations for core point types in the source file.

[mvieth]

Some first review comments, I will probably hove more comments later.
Another question: you claim ~24% reduction in build time, what is that number based on? In the two clang build analyzer reports, total build time of the baseline is 105.3s+197.6s=302.9s (frontend+backend), while afterwards it is 108.7s+198.6s=307.3s (so actually more?). Looking only at the build time of sac_model_ellipse3d.cpp, for the baseline it is 40.3s (again frontend+backend), afterwards it is 37.9s, so about 2.5s or 6% difference.

[AlrIsmail]

Some first review comments, I will probably hove more comments later. Another question: you claim ~24% reduction in build time, what is that number based on? In the two clang build analyzer reports, total build time of the baseline is 105.3s+197.6s=302.9s (frontend+backend), while afterwards it is 108.7s+198.6s=307.3s (so actually more?). Looking only at the build time of sac_model_ellipse3d.cpp, for the baseline it is 40.3s (again frontend+backend), afterwards it is 37.9s, so about 2.5s or 6% difference.

Sorry for the misattribution; the 24% is actually based on wall-clock rebuild time for the pcl_sample_consensus target.

Target: pcl_sample_consensus
Command: time make pcl_sample_consensus -j8
----
Before Changes:
 real    0m23.671s
 user    0m22.813s
 sys     0m1.097s

After Changes:
 real    0m17.473s
 user    0m16.644s
 sys     0m0.996s

While the serial aggregate in the reports is flat, this change reduced LevenbergMarquardt instantiations from 61 down
to 44

[AlrIsmail]

new values after taking into account review comments:
analysis_baseline.txt
analysis_optimized.txt

Metric Category	Pre-Optimization	Post-Optimization	Impact / Gain
Build: Total Time (File)	36,294 ms	34,051 ms	-6.2% Reduction
Template Instantiations	61 instantiations	44 instantiations	-27.8% Reduction
Runtime	1,684.2 μs	1,532.6 μs	+9.0% Speedup

[AlrIsmail]

@mvieth, I've undone any unrelated modifications as requested (except for loops, since I get clang-tidy errors). Could you check?

[azure-pipelines]

Commenter does not have sufficient privileges for PR 6430 in repo PointCloudLibrary/pcl

[mvieth]

Okay, I am happy with the changes. Thanks!

@larshg I am going to leave this PR open a few more days, so you have the opportunity to review again, if you want to. If you don't want to or don't have the time, I would merge by end of next week.

[larshg]

Looks good 👍

[mvieth]

@AlrIsmail BTW, after looking at the clang build analyzer output, I believe modifying this line might also improve build time and run time: the currently used constructor internally calls compute, which requires the instantiation of several expensive templates. If we switch to a different constructor and call computeDirect instead, that may improve things. Though it may be necessary to switch the solver from float to double, to stay accurate enough. See also the documentation of SelfAdjointEigenSolver. I just wanted to mention it, in case you are interested in testing how this change would influence build time and run time.