Add grid() interior aperture sampling (toward robust field/pupil sampling)

[jacobdparker]

Summary

Adds a grid() method to RectangularAperture and CircularSectorAperture — the interior counterpart to wire() (which samples the boundary). It maps a grid of normalized coordinates in [-1, 1] onto points that fill the aperture, with transformation respected. This is the building block for sampling a stop's interior with a grid of rays instead of inferring its extent from the boundary.

This is intentionally a small, self-contained primitive PR. The motivation — and the follow-up that consumes it — is below.

Motivation: the field/pupil sampling is fragile

SequentialSystem._denormalize_grid maps the normalized [-1, 1] field/pupil grid onto each stop's field of view by tracing the stop perimeters (wire()) out to object space, taking the min/max to form a bounding box, and placing the element "center" at (min+max)/2:

min_field = field.min(axis=(axis_field, axis_pupil))
ptp_field = field.ptp(axis=(axis_field, axis_pupil))
result.field = ptp_field * (result.field + 1) / 2 + min_field

(min+max)/2 is a 2-point estimator of the center: it hangs off two extreme rays, which are aberration-smeared marginal rays at the aperture corners. The corner that achieves the extreme hops as the element scans the field, so the assigned center jumps discontinuously.

In a multi-slit spectrograph (MUSE), this puts a spurious ~0.3 mÅ kink into the slit-to-slit separation near specific slits — about ⅓ of the science requirement — that is not physical (it's wavelength- and grating-independent and doesn't shrink with denser [-1, 1] sampling).

The fix this enables

Replace the 2-point estimator with an N-point average: sample the stop interiors with a grid (this PR's grid()), use the same secant/Newton solver already in _calc_rayfunction_stops_only to find the ray through each (field-point, pupil-point) pair, and let the center fall out as the centroid of the grid (= the chief ray through the element center, smooth by construction). Averaging over the whole aperture is insensitive to which corner is extreme.

A prototype that swaps wire → grid on the existing solver and takes the centroid drops the per-slit center error from 4.01 maas → 0.00 maas (the kink vanishes entirely, not just shrinks). The machinery is already there — _calc_rayfunction_stops already back-propagates to object space — so the consumer is a focused change; I've kept it out of this PR because routing the default rayfunction through it is a behavior change worth its own review (and touches value-asserting system tests).

Credit to Roy Smart for the grid-sampling framing.

Tests

test_rectangular_aperture_grid and test_circular_sector_aperture_grid (grid points fill the interior). Full optika/apertures/_apertures_test.py passes (20486 tests), ruff clean.

🤖 Generated with Claude Code

[codecov]

Codecov Report

❌ Patch coverage is 94.44444% with 2 lines in your changes missing coverage. Please review.
✅ Project coverage is 99.31%. Comparing base (a6e6f7f) to head (9b0c605).

Files with missing lines	Patch %	Lines
optika/apertures/_apertures.py	86.66%	2 Missing ⚠️

Additional details and impacted files

@@            Coverage Diff             @@
##             main     #180      +/-   ##
==========================================
- Coverage   99.34%   99.31%   -0.03%     
==========================================
  Files         116      116              
  Lines        5983     6019      +36     
==========================================
+ Hits         5944     5978      +34     
- Misses         39       41       +2     

Flag	Coverage Δ
unittests	99.31% <94.44%> (-0.03%)	⬇️

Flags with carried forward coverage won't be shown. Click here to find out more.

☔ View full report in Codecov by Harness.
📢 Have feedback on the report? Share it here.

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.