equi2cube z-axis orientation: remove the `z_down = not z_down` hack

[haruishi43]

Summary

equi2cube flips the user's z_down argument internally before building the rotation matrices, with a FIXME admitting the reason is unknown:

https://github.com/haruishi43/equilib/blob/master/equilib/equi2cube/numpy.py#L174-L177

# FIXME: not sure why, but z-axis is facing the opposite
# probably I need to change the way I choose the xyz coordinates
# this is a temporary fix for now
z_down = not z_down

The same hack exists in the torch backend (equilib/equi2cube/torch.py, ~L206-L209). It's a band-aid for an un-diagnosed orientation inconsistency, not a real fix.

Likely root cause

equi2cube builds its grid with the opposite latitude convention to the other equi-sampling transforms:

• equi2cube/convert_grid: uj = (0.5 - phi/π)·h - 0.5 → row 0 = +phi (up)
• equi2equi / equi2pers convert_grid: uj = (phi + π/2)·h/π → row 0 = -phi (down)

So equi2cube's xyz/grid construction uses an inverted z relative to the rest of the library, and the z_down = not z_down flip compensates for that mismatch. The proper fix is to make equi2cube's coordinate convention consistent with the other transforms (likely in create_xyz_grid or the convert_grid latitude term) and then delete the flip.

A related # FIXME: needs to test this sits near the output assembly (equilib/equi2cube/numpy.py ~L203), indicating untested paths in the same transform.

Why it matters

This is the one place in a core transform where the behavior is "works but not sure why." Current tests mostly check numpy/torch parity and shapes, not geometric ground truth, so the orientation could be only coincidentally correct for the tested cases (e.g. z_down=True, pitch/roll composition on the up/down faces, all cube_formats).

Proposed work

1. Reproduce with a synthetic equirectangular image whose regions map to distinct, identifiable cube faces; verify each face lands in the geometrically correct place (and across z_down, roll/pitch/yaw, and every cube_format).
2. Fix the coordinate convention at the source so it matches equi2equi/equi2pers, and remove z_down = not z_down from both backends.
3. Add a geometric ground-truth regression test for equi2cube orientation (and resolve the needs to test this FIXME).

Milestone

Deferred from 0.6.0 (pre-existing behavior, not a regression). Target: 0.6.1.

🤖 Filed with Claude Code