Refine reverse NTv2 shifts at the estimated source coordinate

[timothyanderson096-ocdealcheck]

Fixes #169.

Implements the single reverse refinement proposed in the issue. After estimating the source coordinate, the transformation evaluates the grid shifts there and applies those refined shifts to the original target coordinate. An estimate outside the grid raises the existing extent error.

Adds self-contained regression tests covering both interpolation methods, positive and negative shifts, interpolation-cell crossings, repeated round trips and an estimate leaving the grid. Expected coordinates are calculated independently of the forward implementation.

Validation:

• Existing baseline: 74 tests passed.
• New regression cases reproduce the failures before the fix.
• After correction: all 78 tests pass using the repository’s pytest/coverage command.
• Both coordinates reported in NTv2 bicubic and bilinear interpolators are not currently 'reversible'; Add a (single) iteration to find grid shift values in the CRS of the grid #169 were checked with both interpolation methods using the official ICSM GDA94–GDA2020 conformal-and-distortion grid. Previous residuals were approximately 0.14–1.05 mm; all four corrected results exactly matched the original coordinate floats.

One refinement reduces the error; it does not guarantee exact inversion for arbitrary grids. Existing rounding and bicubic boundary behavior remain unchanged, keeping #171 and #170 separate.