Handle None vs. Nan in data association

here is the offending line: https://github.com/borglab/gtsfm/blob/master/gtsfm/data_association/point3d_initializer.py#L187

we pass None as the avg_track_error if poses underconstrained, or cheirality check fails, or inliers underconstrained

but NaN is returned in some other case. so we could just pass np.nan or float("NaN") for each of those

or cast numpy arrays of object type to float32 in metrics, because object types are not supported. But we should not cast everything.

Start at gtsfm/data_association/point3d_initializer.py#L187 and trace how avg_track_error is produced when poses, cheirality, or inliers are underconstrained. Then inspect the metrics code that receives these values and decide which representation is consistently supported. Done means the chosen handling avoids object-typed metric arrays without changing unrelated values.

Written by the indexing model from the issue text.