Add numerical robustness improvements to GJK

[kbayes]

The main changes are:

1. Do a direction fallback when norm becomes too small,
2. Recorrect barycentric coordinates to be exactly a partition of unity, and
3. Some exit condition improvements.

NOTE: One of the tests was actually incorrect and this PR corrects it.

Benchmark Performance (a little regression is expected due to the support fallback):

Benchmark / Task	Before	After	Change (%)
aloha_pot	81.38836	85.13249	+4.60%
aloha_sdf	39.90975	40.09000	+0.45%
aloha_cloth	3839.18790	4036.44060	+5.14%
aloha_clutter[0]	1332.36986	1373.55125	+3.09%
aloha_clutter[1]	63.47256	65.69284	+3.50%
cloth	45.53600	45.92000	+0.84%
franka_emika_panda	0.04578	0.04756	+3.89%
humanoid	0.18512	0.18262	-1.35%
three_humanoids	0.18125	0.18362	+1.31%
myoarm	8.98313	9.50987	+5.86%
unitree_g1_flat	0.20450	0.17312	-15.34%
unitree_g1_hfield	209.68875	214.97988	+2.52%