Pr/lineage clustering

[DonNabla]

Summary

Three-step performance refactor of LineageClustering.build_lineage_for_event:
eliminate redundant np.where rescans, pre-compute the string-field
comparisons as integer codes, and lift the per-particle inner loop into
a single @numba.njit kernel.

The LineageClustering plugin's per-event wall drops by roughly 5×
on realistic γ-background simulations. Bit-identical to the unmodified
plugin (44 M interaction_lineage rows checked across a multi-source
MC sweep, plus a frozen baseline fixture).

Depends on #383 (fix/get-parent-same-time-tag).
This PR is branched off fix/get-parent-same-time-tag so its numba
kernel inlines the spatial tie-break logic introduced there. Once
#383 is merged, this branch rebases trivially onto main.

What changes

Three logically independent commits, each individually bit-identical
to its predecessor:

1. perf(lineage_cluster): eliminate O(N) rescans inside build_lineage_for_event

   • get_parent and get_last_particle_interaction now take a
     precomputed trackid_lookup (built once at the top of
     build_lineage_for_event) and fetch the parent's row indices
     in O(1) instead of re-running np.where(trackid == parent_id)
     for every interaction.
     set tracked inline, dropping a per-particle np.all scan.

2. perf(lineage_cluster): pre-tag string fields with int enums

   • The per-particle classification rules previously did string
     compares on type, parenttype, creaproc, edproc for every
     row. This commit computes an int8 code per row once at the
     top of compute() (unknown strings encode as -1 so they
     never match a named constant), then runs _is_broken_coded
     and _classify_coded against the int codes.
   • Also lifts precompute_ion_AZ out of the hot path (regex is
     applied once per unique type string rather than per row),
     and inlines start_new_lineage / continue_lineage as direct
     field-view writes on the result buffer.

3. perf(lineage_cluster): numba-njit the build_lineage_for_event inner loop

   • The Python per-particle loop is replaced by a single
     @numba.njit kernel _build_lineage_for_event_kernel. The
     wrapper builds a CSR-style trackid lookup (trackid_offsets,
     trackid_indices, trackid_pos, parent_pos) that replaces
     the Python dicts; the kernel walks the per-particle loop
     entirely in compiled code, calling _classify_njit and
     _is_broken_njit (int-only twins of the previous Python
     helpers) for each interaction.
   • main_cluster_type stays on the numpy path because numba
     handles structured-string outputs poorly and it is not in the
     inner-loop hot path.
   • The kernel inlines the parent-finding logic. Following the
     spatial tie-break from Fix get_parent same-time-tag bug in LineageClustering #383, the inlined logic does two
     passes: (1) find t_max = max(t[idx] for idx in slice if t[idx] <= t_i), (2) among entries with t == t_max, pick the
     one closest to the daughter's (x, y, z). Squared distance is
     used to avoid a sqrt — argmin is the same. A nearest-|t - t_i| fallback handles the case where no parent step has
     t <= t_i.

Bit-identity verification

• New unit equivalence test
  tests/test_LineageClustering_equivalence.py asserts
  np.array_equal field-by-field on interaction_lineage against a
  frozen baseline (tests/data/lineage_baseline.npz, 26 306 rows on
  the standard test fixture). Test passes in ~6 s.
• Out-of-tree: interaction_lineage output was also compared field-
  by-field against vanilla 1.6.3 + Fix get_parent same-time-tag bug in LineageClustering #383 on K40, Co60, Th232,
  Ra226, and Pb212 simulated runs — 44 M total rows, all
  identical (and identical between each pair of consecutive perf
  commits, since each step preserves output bit-by-bit).

Performance

Per-source interaction_lineage target wall, before → after, on
realistic Geant4 simulations:

source	vanilla (s)	this PR (s)	speedup
K40	149	30	5.0×
Co60	1920	381	5.0×
Th232	104	27	3.9×
Ra226	140	28	5.0×
Pb212	2271	932	2.4×

(Pb212 has the smallest relative gain because its interaction_lineage
target is dominated by Geant4 ROOT reading + upstream clustering rather
than the per-particle Python loop this PR optimises.)

Notes

• No upstream code outside fuse/plugins/micro_physics/lineage_cluster.py
  is touched (plus the new test file + baseline).
• LineageClustering.__version__ is bumped via Fix get_parent same-time-tag bug in LineageClustering #383's commit; this PR
  does not change it again.
• The baseline npz is small (~25 KB) and committed as a fixture; the
  test gracefully skipTests if absent and can be regenerated by
  python -m tests.test_LineageClustering_equivalence regen.
• First-call JIT compile of _build_lineage_for_event_kernel is
  one-time per process; the cache=True flag persists across
  invocations.