Preserve Fisheye624 Jacobians on the optical axis

[Afloat16]

Fisheye624 projection and unprojection return finite values on the optical axis, but their inactive radial branches divide by zero. Those 0/0 intermediates make the coordinate and camera-parameter gradients NaN. The projection's small-radius branch also replaces the ray direction with (1, 1).

Express projection as a scalar radial scale times the normalized coordinates, with the analytic limit theta_distorted / r -> 1. Protect the inactive denominator in the existing unprojection center branch as well. This restores the center Jacobians: diag(fx/z, fy/z) for projection and diag(1/fx, 1/fy) for unprojection.

Validation:

• Native production imports, including the TorchScript unprojection helper: 13 focused tests passed on CPU Torch 2.14.1; unchanged source: 12 failed, 1 passed. The complete existing tests/cameras directory plus these regressions also passed: 22 passed (existing deprecation warnings).
• Tests cover both 15/16-parameter layouts, float32/float64 center Jacobians, camera-parameter gradients, gradcheck, near-axis direction, distorted batched round trips with gradients, and a noncentral closed-form control.
• Repository Ruff 0.12.2 lint/format checks for changed files, license-header check and git diff --check passed.

The complete own-fork workflow was compared with exact unchanged upstream using identical environments. With latest dependencies, both have the same 19 Open3D type errors, two image-processing failures, missing-EGL collection error and missing-pkg_resources documentation error; baseline 79 passed versus patch 92 passed. License, notebook metadata, Ruff lint/format and tracked-file checks pass in both.

A separate compatibility matrix resolves NumPy 1.26.4, setuptools<81 and Open3D 0.18.0 with the dev/docs dependencies, installs libEGL1 and passes uv pip check. Both now pass Pyright (0 errors) and documentation. Baseline 81 passed versus patch 94 passed, including all 13 new regressions; both retain the same two existing training-test failures in pil_to_numpy / Pillow 12.3.0 encoder setimage. Full suites remain red. Compatibility changes are confined to the own-fork validation workflow.

GPU tests and reconstruction-quality benchmarks were not run. Local focused pytest overrides the parallel/type-checking addopts; the comparative workflows use the repository pytest configuration.