Fix support radius in zonotope.enclosePoints

[suhaslord]

Summary

Fixes the support-radius calculation in zonotope.enclosePoints and prevents a
zero generator from being normalized during residual-cloud compression.

Closes #79.

Problem

The previous implementation used abs(dot(u,r)), where r is the nonnegative
radius vector of an axis-aligned interval and u is the selected direction.
Opposite signs in u can cancel even when the interval has nonzero extent. For
example, u = [-1; 1] / sqrt(2) and r = [1; 1] yield a zero dot product, after
which normalizing the zero generator introduces non-finite values.

Fix

• use dot(abs(u),r), the support radius of the axis-aligned interval in direction u
• return from compression when the generated direction has zero norm
• add regression coverage for a degenerate point cloud and a sign-cancelling cloud

Validation

• verified the sign-cancellation reproduction independently
• verified that dot(abs(u),r) equals the maximum absolute projection over the interval vertices
• checked the final diff against upstream master; only the implementation and focused unit test change
• ran git diff --check successfully
• MATLAB and Octave are not installed locally, so the MATLAB unit test was not executed locally

[toladnertum]

Dear @lili1039 and @suhaslord,

Thank you so much for bringing this up (#79) and providing a solution alongside it here!

We will investigate the bug a bit further and run it through our internal CI/CD pipeline,
and hope to push the fix out soon with the next release of CORA.

Best,
Tobias