fix findTriTriDistance in case of finite upDistLimitSq

[Fedr]

findTriTriDistance could return a wrong (overestimated) distance when a non-default upDistLimitSq was given.

In the edge-pair loop, sd.dir is the vector between the closest points of the two edges and is not unit length. The projections p, s, t are dot products with sd.dir, so they carry an extra factor of |sd.dir|, and sqr( p - s + t ) overestimates the actual squared distance between the separating planes by |sd.dir|². This inflated value was compared against upDistLimitSq and, on early exit, stored in res.distSq — so with a finite limit the function could stop early and report a distance larger than the true one (and larger than what the same call without a limit returns).

The fix divides the squared plane distance by sqr( sd.dir ) (guarding against a zero vector), making the early-exit value a true squared distance.

Added a regression test with two triangles ~277.5 apart: the distance computed with upDistLimitSq slightly above the true squared distance must match the unlimited result; before the fix the early exit returned an inflated value.