sample_consensus: reduce SampleConsensusModelEllipse3D template instantiations

Move the Levenberg-Marquardt optimization logic into a non-templated
pcl::internal::optimizeModelCoefficientsEllipse3D function in the
source file, reducing redundant template instantiations from 61 to 44.
This results in a ~7% reduction in build time and 10% runtime speedup
for the pcl_sample_consensus library.

Move the get_ellipse_point, dvec2ellipse, and golden_section_search
math helpers into pcl::internal in the header to allow shared access
without exposing them as class members.

Additionally:
- Fix projectPoints to write to projected_points[inlier] instead of
  projected_points[i++] when copy_data_fields is true, consistent with
  other SAC models
- Fix modernize-return-braced-init-list lint warning

Author: AlrIsmail