Slice-package splitting requires VisuCoreSlicePacksSlices, which PV5.1 never writes (spec 7.10) #169

From §6.2 of the format-conformance review against
FILE_FORMAT.md.

The gap

§7.10 says the slice-package parameters are optional and were introduced in
PV6.0.1 — PV5.1 defines none of them — and may be absent even in PV6/PV360
datasets.

After #156 and #165, Dataset.slice_packages_index() copes with that: it reads
VisuCoreSlicePacksSlices when present, honours the -1 fg_index sentinel,
and otherwise groups contiguous frames that share an orientation, so
affine_of_package(i) works on a PV5.1 multi-package localizer. The splitter
does not:

>>> SlicePackageSplitter().split(pv51_localizer)
MissingProperty: Parameter VisuCoreSlicePacksSlices not found

So get_slice_packages() / Dataset.slice_packages — the documented way to get
per-package data and geometry — is unavailable for exactly the datasets whose
geometry most needs splitting. num_slice_packages reports 1 for them, so
nothing warns either.

Proposal

1. Have SlicePackageSplitter and num_slice_packages use
   slice_packages_index() as their single source of package boundaries, so
   the PV5.1 inference and the -1 sentinel apply everywhere.
2. Keep the per-package VisuCoreSlicePacksSlices/SliceDist rewriting for
   datasets that have those parameters, and synthesise a one-package descriptor
   for those that do not, so a written package is still a valid dataset.
3. Test on the shape the review found: three corpus PV5.1 localizers with two
   orientations and no slice-package parameters.

Caveat worth recording: the inference groups contiguous frames sharing an
orientation, so two adjacent packages with the same orientation would merge.
No corpus dataset has that shape, and PV6+ datasets never rely on the inference.