Fix MATLAB wrapper virtual class downcasting for namespaced types

[thatdudegrantt]

Problem
gtsam.noiseModel.Gaussian.Information(Q) with a scalar/diagonal matrix correctly downcasts to Isotropic in C++, but the MATLAB wrapper always returned a gtsam.noiseModel.Gaussian object. This caused JacobianFactor and similar functions that require a diagonal noise model to fail.

Root cause (two bugs)
_collector_return and wrap_collector_function_return_types always passed isVirtual=false to wrap_shared_ptr, bypassing the RTTI registry entirely even for classes declared virtual in the .i file.
generate_preamble stored flat concatenated names (e.g. gtsamnoiseModelIsotropic) in the RTTI registry instead of dot-separated MATLAB package names (e.g. gtsam.noiseModel.Isotropic), causing mexCallMATLAB to fail for any namespaced virtual class.

Fix

Check cls.is_virtual against self.classes in both return codegen functions and pass true to wrap_shared_ptr for virtual classes.
Use _format_class_name(cls, separator='.') for RTTI registry entries.

Testing
Verified end-to-end on Windows x64, MSVC 19.51, MATLAB R2026a:

nmQ = gtsam.noiseModel.Gaussian.Information(2);
class(nmQ)  % returns gtsam.noiseModel.Isotropic
gtsam.JacobianFactor(0, 0, 0, nmQ)  % succeeds

All 10 existing wrap Python tests pass. Added virtual base/derived fixture to tests/fixtures/inheritance.i to cover the isVirtual=true codegen case.
Fixes borglab/gtsam#605

[thatdudegrantt]

CI is passing. Verified end-to-end on Windows x64, MSVC 19.51, MATLAB R2026a. noiseModel.Gaussian.Information(2) now returns gtsam.noiseModel.Isotropic and gtsam.JacobianFactor(0, 0, 0, nmQ) succeeds. Ready for review, thanks in advance!

[ProfFan]

Cool I'll review today