Fix torch.atan2 for y == 0 with negative x

[LeSingh1]

Problem

torch.atan2 returns the wrong quadrant in two cases.

The rotation by π for x < 0 is gated on strictly y > 0, so atan2(0, x) for x < 0 returns 0 instead of π. Separately, the "avoid divide-by-zero" shift adds a fixed +2e-8, which flips the sign of any x in (-1e-8, 0) — moving atan(y / x_safe) into the other half plane while the quadrant coefficients still correct for x < 0, giving another ±π error.

y = [0, 0,  1, -1]
x = [-1, -3, -1e-9, -1e-9]

torch : [ 3.14159,  3.14159,  1.57080, -1.57080]
before: [ 0.00000,  0.00000,  4.71239, -4.71239]
after : [ 3.14159,  3.14159,  1.57080, -1.57080]

Fix

Use greater_equal(y, 0) for the x < 0 rotation, and make the safe shift follow sign(x).

The x == 0 branches keep strict > / <, so atan2(0, 0) == 0 is preserved to match PyTorch.

Testing

Added test_atan2_y0_xnegative and test_atan2_x_tiny_negative. Before: 6 failed, 2 skipped. After: 46 passed, 2 skipped, and the full TestAtan2 class passes 84.

A 2000-element random sweep including zeros and tiny negatives shows 0 mismatches against eager PyTorch, max error 2.4e-7.

The existing TestAtan2 covers random inputs, x == 0, and y == 0 & x == 0 — the y == 0, x < 0 quadrant was never exercised.

[TobyRoseman]

CI: https://gitlab.com/coremltools1/coremltools/-/pipelines/2731449720