reproject: unit-aware bounds_policy="auto" blow-up detection

[brendancol]

Summary

Closes #2582. The bounds_policy="auto" heuristic compared source span (in source CRS units) to output span (in target CRS units), so the > 50x ratio fired on almost every geographic-to-projected reprojection and silently cropped valid edge data via the 2/98 percentile fallback. For a (-10, -10, 10, 10) EPSG:4326 bbox going to EPSG:3857, that was 70-106 km per side.

Replaced the unit-mismatched ratio with two unit-agnostic signals:

• Magnitude ratio: max absolute projected coordinate / median absolute projected coordinate. Benign reprojections stay near 1-2. Real singularities push past 10 (Mercator near the poles after clamp) or far higher (polar-stereographic on the opposite pole projects to ~1e23).
• Non-finite fraction: share of raw (unclamped) edge samples that produce inf/nan in the target CRS. Catches singularities that the auto clamp would otherwise hide.

Verification

The reproduction case from the issue:

	Before	After
auto vs raw x crop	106 km	0 m
auto vs raw y crop	70 km	0 m

Global EPSG:4326 -> EPSG:3857 still trips the percentile fallback. The pathological EPSG:4326 -> EPSG:3413 case (south pole into a north polar stereographic) still trips.

Backend coverage

• numpy: yes
• cupy: yes (heuristic runs on a small numpy probe, no GPU-specific path)
• dask + numpy: yes (regression test added)
• dask + cupy: yes (same as cupy)

Test plan

• Regression test: 4326 -> 3857 on (-10,-10,10,10) under auto matches raw within one pixel.
• Benign 4326 -> 3857 emits no bounds_policy warning under auto.
• Pathological 4326 -> 3413 still trips the auto fallback and warning.
• Existing 13 TestBoundsPolicy tests pass.
• Full reproject suite (368 tests) passes.

[brendancol]

PR Review

Blockers

None.

Suggestions

• xrspatial/reproject/_grid.py:271-294. The non-finite probe duplicates the edge-sampling pattern from lines 215-226, just with the unclamped source coordinates. A small helper would keep the two samplers in sync if n_edge is ever tuned. Optional given the surgical-changes rule, but flagging it.

Nits

• xrspatial/reproject/_grid.py:299. The nonfinite_frac > 0.05 threshold is reasonable but unexplained. A one-line comment ("5% means edges hit a singularity; benign cases yield 0%") would help future maintainers understand the choice.
• xrspatial/reproject/_grid.py:295-298. The magnitude-threshold comment mentions "ratio ~1e16" for the opposite-pole case. Worth also noting that the post-clamp Mercator-global case has a ratio around 5 (caught only by the non-finite probe), so the reader sees why both signals are needed.

What looks good

• The bug is reproducible before the fix: 106 km x crop, 70 km y crop on (-10, -10, 10, 10) into EPSG:3857.
• Two-signal design cleanly separates the pathological cases. The magnitude ratio catches finite-but-huge singularities like polar-stereographic on the opposite pole. The non-finite probe catches Mercator at the poles after clamp would otherwise have hidden the issue.
• Tests cover the regression case, the silent-no-warning case, the polar-stereographic pathological case, and a dask-backed regression.
• CHANGELOG and docs both updated.
• All 13 existing TestBoundsPolicy tests pass; full reproject suite of 354 tests passes.
• The non-finite probe is gated on bounds_policy == "auto" and source_crs.is_geographic so the extra pyproj call only fires when needed.

Checklist

• Algorithm matches the issue's suggested approach
• Backends consistent
• NaN handling correct
• Edge cases covered
• Dask test added
• No premature materialization
• Benchmark not required
• README not applicable
• Docstrings updated

[brendancol]

Follow-up review

Addressed all three findings from the previous review.

Fixed

• Suggestion (_grid.py:271-294): extracted _edge_samples() helper; both the main edge sampler and the non-finite probe now share the same edge construction.
• Nit (_grid.py:299): added a comment explaining why 5% is the non-finite threshold (benign cases yield 0%, real singularities yield 15%+, the floor absorbs numerical noise).
• Nit (_grid.py:295-298): expanded the magnitude-threshold comment so the reader sees why both signals are necessary. The post-clamp global Mercator case has ratio ~5, well under 10x; the non-finite probe is what catches that case.

All 17 TestBoundsPolicy tests still pass; full reproject suite of 358 tests passes.