Memory-efficient TCA

[quannfa]

This PR improves the memory efficiency of TCA on large datasets and fixes a normalization edge case that could produce NaN values. The changes are focused on the TCA fit implementation.

Background

In the original implementation, the primal mode explicitly constructs n×n M and H matrices. With large sample sizes (for example, around 56k samples), this leads to very high memory usage and can cause OOM failures.
Also, column normalization did not handle zero-norm columns, which could result in NaN values.

What Changed

Reworked the primal-kernel computation path to avoid explicitly building n×n M/H matrices.
Replaced matrix products with low-rank equivalent forms:
K·M·Kᵀ is computed via an outer product based on Xe.
K·H·Kᵀ is computed as XXᵀ minus a rank-1 centering update.
Added safe normalization for both input X and projected Z, normalizing only columns with norm > 0.
Kept the non-primal branches (linear/rbf) as fallback with the original logic.

Expected Impact

Significantly lower memory usage in primal mode, from O(n²) to approximately O(m·n + m²).
Better scalability and reduced OOM risk on large datasets.
Improved numerical stability by preventing NaN generation from zero columns.

Compatibility

No public API changes.
Primal branch internal computation path changed but remains mathematically equivalent.
Linear/RBF branches preserve existing behavior.

Validation Notes

The commit modifies only the TCA implementation file.
Full end-to-end regression/training was not run in this step; it is recommended to run a full pipeline check on representative datasets.