perf: avoid post-pass column reconstruction in NaN preprocessing

[CyrilJl]

Summary

This PR optimizes the preprocessing step used to build the NaN coordinate mapping in OptiMask.

Instead of reconstructing cols_with_nan at the end with np.flatnonzero(...) and an additional reordering pass, the branch records NaN column indices directly
during the main scan. That removes extra array work after preprocessing while preserving the existing output shape and semantics.

What Changed

• Add a dedicated cols_with_nan buffer during preprocessing
• Fill that buffer the first time a NaN is seen in a column
• Slice the populated portion at the end instead of rebuilding the column list from cols_index_mapper

Why

The previous implementation did the column discovery twice:

• once implicitly while scanning values
• once again at the end to rebuild ordered column indices

This change keeps the same behavior but avoids that post-processing step, which should reduce preprocessing overhead on inputs with missing values.

Notes

• Functional change is limited to optimask/_optimask.py