fix: PointSolver zero-image and coarse-precision crashes, plus two consumer gaps

[Jammy2211]

Phase 1 of the @rhayes777 API-audit campaign (epic PyAutoArray#415, task PyAutoArray#416).
Fixes both PointSolver crashes reported in #531, plus two consumer gaps an independent
review surfaced.

Pairs with PyAutoArray#417 — merge that first (this branch's suite runs against it).

The two reported crashes

Both are inputs a user reaches on purpose.

1. Source outside the tiled region → numpy.exceptions.AxisError

No triangle traces to the coordinate, so filtered_means has shape (0, 2); the
[pair for pair in ...] comprehension then yields [], Grid2DIrregular([]).array is a bare
list, and the axis=1 reductions blew up.

Zero images is a legitimate answer, so solve() now returns a correctly-shaped empty grid and
logs why. (The report predicted a single image here. The tiling genuinely finds none — the grid
spans ±2″ and the source sits at (5,5) — so returning an empty grid is the honest answer rather
than fabricating one.)

2. Loose pixel_scale_precision → IndexError: list index out of range

n_steps is ceil(log2(scale / pixel_scale_precision)), which goes negative once the
requested precision is coarser than the initial triangle scale. The existing guard tested == 0,
so negatives slipped through, range(-1) yielded no steps, and steps[-1] raised.

Now tests <= 0 and reports pixel_scale_precision, the triangle scale, and a workable value.
Deliberately not clamped to 1 step — that would silently solve at a precision the caller
never asked for.

Two consumer gaps this exposed

Both predate this branch and were unreachable while solve() crashed first. Found by an
independent Codex review of the first commit.

FitPositionsImagePairAll.chi_squared returned NaN with no model positions. n_permutations
is 0, so -log(0) is +inf while the permutation sum is -inf. fitness.py:262 converts a NaN
log-likelihood into resample_figure_of_merit — so the model was silently resampled instead of
scored, precisely what no_image_residual exists to prevent. FitPositionsImagePair and
FitPositionsImagePairRepeat both already applied that floor; PairAll was the one sibling that
did not. It now returns the same value they do (asserted equal, not merely finite).

Selected with xp.where, not a Python if — the model-position count is a traced value under
jax.jit. The log is taken on a clamped count so the NaN never forms in the discarded branch.

Result.image_plane_multiple_image_positions recovered only at exactly one image. Zero fell
through as an empty grid and SourceMaxSeparation reduced over it (max() on an empty sequence).
Zero is the case that most needs the inward-walk recovery: == 1 → <= 1.

API Changes

Behavioural change, no signature change.

• PointSolver.solve(...) returns an empty (0, 2) Grid2DIrregular where it previously
  raised AxisError. Callers already handling a variable image count need no change; anything
  assuming ≥1 image now sees length 0 rather than an exception. Verified .shape[0], len(),
  .in_list and re-wrapping all behave.
• solve(...) raises a clearer ValueError (was IndexError) when pixel_scale_precision
  exceeds the triangle scale. The guard lives on AbstractSolver, so ShapeSolver gains the same
  message.
• FitPositionsImagePairAll gains no_image_residual = 1.0e4, matching its siblings. Fits that
  previously produced NaN now produce a finite, loudly-bad chi-squared.
• A WARNING is logged whenever solve() returns empty, naming magnification_threshold.

Downstream impact: none expected — every changed path previously raised or produced NaN.

Testing

• 11 new tests: both reported crashes, both routes to an empty result (outside-region and
  all-candidates-filtered), a no-warning control, the 4-image control, and the PairAll floor
  asserted equal to both siblings.
• PyAutoLens 488 passed · PyAutoArray 930 · PyAutoGalaxy 1009.
• Controls: PointSolver at precision 0.001 still returns its 4 images.

Reported by @rhayes777. Closes #531.