Fix LAEA inverse rq normalization and authalic inverse series

[brendancol]

Closes #3274

• _laea_inv_point (oblique/equatorial) stripped the rq factor in its input normalization and then divided by rq again in the angular distance, inflating distances by ~0.11% of the distance from the projection origin. Keep rq in rho so sce = 2*asin(0.5*rho/rq) divides by it exactly once, matching PROJ's laea.cpp. Same change in the CUDA kernel _d_laea_inv.
• _authalic_apa coefficients did not invert _authalic_q (wrong at leading order in the second term, 17/360 vs 23/360). Replaced with PROJ's pj_authset / Snyder eq. 3-18 coefficients. This feeds the AEA, CEA, and LAEA inverses on CPU and CUDA; array shape and kernel signatures are unchanged (terms 4-6 are now zero).

Parity vs pyproj after the fix: EPSG:3035 inverse 1.0 mm (was 2.6 km), EPSG:6933 inverse 1.6 mm (was 4.8 m), WGS84 AEA inverse 1.5 mm (was ~4.9 m). Forward kernels were already at the 1e-5 m level and are unchanged. The residual ~2 m for EPSG:5070 is pyproj's WGS84->NAD83 step, which the fast path skips by design.

Backends: numpy and dask+numpy share the CPU kernels; cupy and dask+cupy use the fixed CUDA kernel. CUDA tests ran on a real GPU.

Test plan:

• New test_reproject_inverse_kernels_3274.py: series round-trip vs _authalic_q, coefficient pin, inverse parity vs pyproj for LAEA/CEA/AEA (oblique and polar), forward regression pin, end-to-end fast path vs transform_precision=0, CUDA-vs-pyproj and CUDA-vs-CPU parity
• Full reproject suite: 607 passed

[brendancol]

PR Review: Fix LAEA inverse rq normalization and authalic inverse series

Blockers (must fix before merge)

None.

Suggestions (should fix, not blocking)

• _projections.py _laea_inv_point: sce = 2.0 * math.asin(0.5 * rho / rq) has no clamp on the asin argument, unlike the ratio clamps elsewhere in the same function. With the corrected (larger by 1/rq) rho, inputs beyond the projection disc reach the asin domain edge slightly sooner and produce NaN instead of a clamped pole value. NaN for out-of-disc input is defensible (PROJ errors there) and the downstream finite-filters handle it, but a one-line clamp would match the file's own conventions. Same spot in _d_laea_inv (_projections_cuda.py).
• test_reproject_inverse_kernels_3274.py covers OBLIQ (3035) and N-polar LAEA but not the EQUIT branch (lat_0=0), which shares the fixed code path. I ran an EQUIT parity check locally (+proj=laea +lat_0=0 +lon_0=20, 500 points): max error 1.6 mm, so the fix holds there too; a pinned test would keep it that way.

Nits (optional improvements)

• _authalic_inv still evaluates five series terms although terms 4-5 are now zero. The docstring explains the signature-stability tradeoff, so this is fine to leave; mentioning the dead terms could alternatively be dropped from the loop in a later cleanup.

What looks good

• The rq fix matches PROJ's laea.cpp normalization exactly (x /= dd; y *= dd with rho keeping rq), and the comment documents why the factor must stay in rho.
• The replacement coefficients are pinned against the exact _authalic_q numeric inverse (round-trip test at <5e-9 rad) rather than against pyproj alone, so the test distinguishes a series bug from a datum-step difference.
• The AEA test deliberately uses a WGS84 proj-string CRS instead of EPSG:5070, with a comment explaining the ~1-2 m NAD83 step that pyproj inserts and the fast path skips. That avoids a flaky tolerance.
• CUDA kernels updated in lockstep and verified two ways: against pyproj per pixel and against the CPU fast path at 1e-9.
• Kernel signatures and the _APA array shape are unchanged, so no call-site churn on either backend.

Checklist

• Algorithm matches reference (PROJ laea.cpp, pj_authset/Snyder 3-18)
• All implemented backends produce consistent results (CPU/CUDA parity test)
• NaN handling unchanged; finite-filtering downstream unaffected
• Edge cases: polar branch pinned, forward path pinned, end-to-end vs exact path
• EQUIT LAEA branch untested (verified manually, see suggestion)
• No materialization or copy concerns (scalar kernel math only)
• Benchmark not needed (no perf-relevant change)
• README matrix not applicable (no new function)
• Docstrings updated with measured error bounds and issue reference

[brendancol]

Follow-up review after the suggestion commit: both suggestions are addressed. The asin argument is now clamped in both the CPU kernel and the CUDA twin with a comment tying it to the existing ratio clamps, and the EQUIT-mode parity test pins the lat_0=0 branch at the same 5 cm tolerance (measured 1.6 mm). The nit about the two zero series terms stays as documented signature-stability. 11 regression tests and the full 460-test reproject suite pass. No remaining findings.