Bump version: 1.1.4 -> 1.2.0

[gouwens]

Prepares the v1.2.0 release. Version is bumped in pyproject.toml (the only place it
lives, since .bumpversion.cfg was removed in the uv conversion) and uv.lock is
relocked to match.

A minor bump rather than a patch: several fixes since v1.1.4 change results, not just
internals, so downstream analysis code should not pick this up silently from a ~=1.1.0
pin.

Results that change

• Right-hemisphere reflection was one voxel out. IsocortexCoordinateProjector now
  reflects the micron coordinate and takes the voxel that mirrored point falls in, so
  the voxel that picks the streamline and the coordinate that measures depth along it
  always agree (IsocortexCoordinateProjector reflects voxels and coordinates by conventions that differ by one voxel #27). angle.find_closest_streamline mirrors correctly at the midline
  in both directions (find_closest_streamline accepts resolution but does not use it to voxelise #23, find_closest_streamline mirrors hemispheres one voxel out, and skips the midline plane #33).
• project_volume no longer corrupts the caller's array. The max/min sentinel
  is written into the gathered copy instead of volume.flat[0] (project_volume silently corrupts the caller's volume #20).
• Path-ordered data projects from path_ordering, not the 3-D volume lookup (project_path_ordered_data cannot run against the current-generation assets #22).

Silent-wrong behaviour now raises

• Invalid kind or scale raises ValueError instead of returning zeros or hitting
  UnboundLocalError (Unrecognised arguments fail silently or obscurely #21).
• An unordered lookup column raises ValueError naming the first out-of-order entry,
  rather than letting searchsorted return wrong voxels.

Other fixes

• SWC files parse with any whitespace delimiter (ccf_streamlines.morphology issue with load_swc_as_dataframe with different delimitier when processed with MorphIO v3.3.3 #8).
• BoundaryFinder keeps an absent region's empty boundary 2-D so reflection is a
  no-op (region_boundaries crashes when reflecting a region absent from the atlas #25); empty masks and an unraw regex separator fixed.
• project_coordinates names projection_file when it needs one (IsocortexCoordinateProjector without projection_file fails obscurely on first use #24).
• The isocortex layer key list has a single definition shared by both projectors and
  metrics (The layer key list has three independent definitions #26).
• An empty nearest-streamline search is handled instead of raising.

Performance

• _matching_voxel_indices resolves membership from the binary search instead of a
  full-column scan, and the ordering check is memoised per array — this was 12.5 s per
  coordinate against the 61.9M-row closest_surface_voxel_lookup.h5.

Packaging and tests

• Project converted to uv (uv_build backend); packaging regressions from that
  conversion fixed, documentation URL added to project metadata.
• Test suite added: mini-CCF fixtures build a ~20 KB structurally faithful atlas, so
  the projector classes are testable without the multi-hundred-MB assets. CI gates PRs
  on x86-64 and ARM64 across Python 3.9 and 3.13. An opt-in real_data tier asserts
  the fixture's format claims against the real files.

Verification

• uv run pytest: 366 passed, 36 skipped (the skips are the opt-in real_data tier).
• python -m build produces ccf_streamlines-1.2.0.tar.gz and the matching wheel;
  installed into a clean environment, all nine modules import and __version__
  reports 1.2.0.

Merging this and publishing a v1.2.0 GitHub release triggers python-publish.yml,
which builds and uploads to PyPI via trusted publishing.

🤖 Generated with Claude Code