Fix KDE all-zero output with descending-coordinate templates

[brendancol]

Fixes #1198

Proposed changes

• Fix bounding-box index calculation in _kde_cpu and _line_density_cpu to handle negative dx/dy (descending coordinate order). The old code divided by dx/dy to get pixel indices, but negative spacing flipped lo and hi so the inner loops never executed, producing all-zero output.
• The fix computes both candidate indices and uses min/max to always get lo <= hi, regardless of spacing sign.
• Add 10 regression tests covering descending y, descending x, both descending, all kernel types, and line_density.

Test plan

• All 35 existing test_kde.py tests still pass
• 10 new TestDescendingCoordinates tests pass, covering:
  • Single-point KDE with descending-y template (nonzero check)
  • Descending-y vs ascending-y match (row-flipped, exact for single point)
  • Descending-x vs ascending-x match (col-flipped)
  • Both-descending vs ascending match
  • All three kernels with descending-y and multiple points
  • line_density with descending-y (quartic exact, gaussian within atol=1e-4)
  • line_density with descending-x
• Dask parity tests pass (Dask tiling inherits the CPU kernel fix)
• CuPy parity tests pass (CUDA kernel was never affected since it has no bounding box)