Propagate product metric parallel transport to factors - #2131

Fixes #2130, reported by @luisfpereira.

ProductRiemannianMetric inherits an unimplemented parallel_transport, so PolyHyperbolicCholeskyMetric works for 2×2 correlation matrices but fails for 3×3 and larger ones. The product connection splits across factors; delegate transport to those metrics using the existing argument projection and output packing. Factors without transport continue to raise their own unsupported-operation error.

Tests cover direction/end-point paths, flat/stacked representations, batching, factor results, tangency and norm preservation. A 3×3 correlation regression checks both paths and reverse transport.

Verification and unmet prerequisites

This draft does not meet the contributing guide's all-tests-pass prerequisite.

Exact head 22959f5045dc87b13ef4e57267d454e58598912e, Python 3.12.14 with supported dependency bounds, CPU:

• Five new regressions pass on NumPy, Autograd and PyTorch and fail on pristine 9941df0a3 (NumPy). Four affected native modules passed 2641 tests per backend, with 49 skips.
• Full pytest --cov=geomstats tests/ on NumPy, including slow/redundant/script nodes: 17284 passed, 26 failed, 556 skipped, 61 xfailed, 212 xpassed; every one of the 18139 collected nodes is accounted for. Six failures are LieCholeskyMetric assertions. Twenty are notebook kernel startup blockers: ZeroMQ reports Operation not permitted, followed by Kernel died before replying to kernel_info; their notebook cells were not executed. A supported escalated kernel smoke has the same failure.
• The guide explicitly requires ≥90% coverage for new code/features. All four executable statements in the sole new production method have 100% line coverage in the full NumPy report and focused reports on all three backends. Unfiltered whole-project NumPy line coverage is 84.223%, reported separately.
• Full Autograd execution was killed (exit 137) before completion: 6225 passed, 1 failed, 577 skipped, 31 xfailed, 150 xpassed, 1 active node interrupted and 11190 collected nodes unrun. No final full-suite coverage report was produced. Full PyTorch execution has not started and is held for resource coordination. Earlier non-slow/non-redundant runs had separate OpenHemisphere exp/log, Autograd where dtype (Fix where for autograd #2118/Fix Autograd where output type #2129), and TorchLBFGS failures on unchanged paths. Baseline evidence does not waive the full-test prerequisite.
• Changed files are formatted; Ruff has the same four pre-existing diagnostics as pristine source. git diff --check and Sphinx -W rendering of the new method pass.

Full-project documentation, GPU and the CI Python-version matrix were not run locally.

Checklist

• Clear title, vectorized code and appropriate unit tests
• New code follows the project style and API
• New method is documented and its docstring renders
• New-code executable line coverage is ≥90%
• Relevant issue is linked
• All repository tests pass on applicable backends

Current CI gate

DeepSource reports failure with documentation coverage at 58.9%, below the repository threshold, and warns that its branch baseline may be inaccurate. Its nine “introduced” findings all refer to existing definitions that are AST-identical to pristine 9941df0a3; none points to the new transport or tests. The run needs maintainer baseline/gate review. GitHub Actions are also awaiting approval.