fix: make circular ell_comps image gradients finite

[Jammy2211]

Summary

• evaluate Sersic eccentric radii directly from Cartesian ell_comps, avoiding the undefined polar angle derivative at (0, 0)
• preserve the existing fac <= 0.999 clamp and transformed-grid compatibility
• add NumPy parity coverage for circular, anisotropic, rotated, and near-clamped profiles

Closes #570.

Why

The profiling hazard likelihood.imaging-sersic.ell-comps-origin-nonfinite-gradient showed a finite off-centre Sersic image and complete FitImaging likelihood at the circular point, but a [NaN, NaN] JAX gradient. The sampled parameters are Cartesian ell_comps_0/1; converting them to axis ratio and angle introduces an avoidable polar singularity.

The new radius identity remains in Cartesian coordinates:

r_ecc^2 =
  ((1 + f^2 - 2 e_x) x^2 - 4 e_y x y + (1 + f^2 + 2 e_x) y^2)
  / (1 - f^2)

where f^2=e_y^2+e_x^2 after the existing magnitude clamp.

Validation

• focused current Sersic tests: 8 passed
• NumPy direct/legacy image parity: maximum absolute difference 1.46e-15 across circular, anisotropic, rotated, and near-clamped cases
• downstream JAX jit(grad) image probe at ell_comps=(0, 0): finite; central differences at 1e-4, 1e-5, and 1e-6 agree
• downstream full FitImaging likelihood: value unchanged at -15.477240141252718; origin gradient now finite at [1.1569945714414, -0.5449763942867472]; all ±1e-8 and ±1e-6 neighbourhood gradients finite
• formatting check and compile check passed
• JAX validation is downstream rather than in PyAutoGalaxy unit tests, per repository policy

No public API, prior, configuration, or release change.