Handle strided arrays in Sinter decoding

[danielgaskins]

decode_shots_bit_packed accepted non-contiguous NumPy views, but treated the bytes within each shot as contiguous. For a view such as shots[:, ::2], this could read the wrong bytes and decode a different set of detectors.

Use pybind11's two-dimensional array accessor so both strides are respected. Add a regression test where the second byte of each shot is non-contiguous.

Tests:

• bazel test --jobs=1 src/... //docs:tutorial_jupytext_sync_test
• clang-format-14 --dry-run --Werror src/tesseract_sinter_compat.pybind.h

[danielgaskins]

Yes, this also handles Fortran-contiguous arrays. unchecked<2>() uses both NumPy strides for each (shot, byte) access. I checked a two-byte Fortran-order input locally, and it produced the expected predictions.

For C-contiguous input, the accessor does a little more index calculation than the previous raw pointer access. I have not benchmarked the two implementations, so I do not want to claim zero overhead. The decoding work is likely to dominate, but a contiguous fast path could be added later if profiling shows this matters.

[danielgaskins]

Current main now handles this in the shared decode_sinter_shots_bit_packed helper added by #256, with coverage for row-strided, column-strided, and Fortran-ordered inputs. That covers the original bug more broadly than this branch, so I’m closing this PR as superseded. Thanks for the reviews and branch updates.