Fix heightfield contact correction for transformed heightfields

First, thank you to the developers for your hard work on Coal.

I noticed an issue in the witness-point calculations for heightfields and believe I have narrowed down a fix. This may also be related to #654.

Problem

The heightfield-specific binCorrection() path mixed heightfield-local bin geometry with world-space contact and support points. Existing identity-pose tests did not expose the mismatch. I originally encountered this on version 3.0.1 and reproduced it on the devel branch as well.

For example, colliding a sphere near the center of a translated flat heightfield returned a witness in its local coordinates, rather than world coordinates:

actual:   [0.09, -0.11, 0]
expected: [2.05, -2.07, 0]

In this failure path the penetration depth can still be correct, which made the problem easy to miss. Rotated heightfields could also return an incorrectly transformed normal.

Reproduction

The new test case in hfields.cpp compares a flat HeightField<OBBRSS> with an analytically equivalent box under identity, translated, and translated-and-rotated poses. It checks penetration, both witnesses, and the contact normal using the default CollisionRequest settings.

The test fails on devel before this change and passes with the fix applied.

Fix

This change performs face selection, support mapping, and projection in the heightfield-local frame. It then transforms the corrected witnesses and normal back to world space. The public heightfield API is unchanged.