aurora: read the MAP relative-velocity direction byte

[deisenstein]

We need to update the reader to parse the new velocity direction information in both the maplog and the particle subsample.

From Claude:

Abacus now stores the DIRECTION of a MAP's velocity relative to the mean local DM alongside the magnitude it already kept, as an Nside=4 NEST healpix pixel index (192 pixels, one byte). Surface it as a plain uint8 rather than expanding to three floats, to keep the post-processing footprint at 1 byte/particle; callers index the lookup table only when they actually want vectors:

from abacusnbody.data.output_particle import rel_vel_healpix_lookup
v3 = t['rel_vel'][:, None] * rel_vel_healpix_lookup[t['rel_vel_healpix']]

output_particle: the MAP flag is now the pid_k band (pid_k & 0xff00) == 0xff00 rather than the single value 0xffff, with the low byte carrying the direction. LAYOUT is bumped to output_particle_24b accordingly -- an older reader pointed at a new file would have decoded most MAPs as DM particles with no error at all.

maplog: the multiplicity word is now a count in the low 24 bits with the direction in the top byte, so mult is masked to 0x00FFFFFF (unconditionally, so every modality decodes the same way). Adds a LAYOUT of mapnode_24b, which read_abacus now validates; files written before the split carry no MapLogLayout key at all, so an absent key is accepted and a mismatched one is not.

Renames maplog's vel_rel to rel_vel. This is a breaking rename, done deliberately: the two modules had different names for the same physical quantity, and propagating that into a second field name would have entrenched it.

The 192 unit vectors live in _healpix_nside4.py, generated from the same source as the C header and re-exported from both readers. Abacus also writes the table into every ASDF header as the 'rel_vel_healpix_lookup' ndarray, so files remain self-describing and no healpix library is needed; the module-level copy just saves opening one.

Adds tests/test_aurora_readers.py, the first tests of any kind for the Aurora readers. Rows are built by hand rather than from reference files: these formats are pure bit layouts, so a hand-built row states the layout in a form a reviewer can check against the C++ by eye. Covers both shared-word splits at their boundaries (0xfeff is the largest DM pid_k; a full-24-bit count beside a nonzero direction byte is what a stale reader gets wrong), all four MapNode modalities, and the invariant that the lookup table inverts the vec2pix_nest64 that Abacus uses to produce the indices.