perf(pmt_afterpulses): two-pass numba scheme for photon_afterpulse

[DonNabla]

Summary

Replaces the per-element Python orchestration in
PMTAfterPulses.photon_afterpulse with a two-pass @numba.njit scheme.
Bit-identical to the unmodified plugin (verified field-by-field on
~30 M pmt_afterpulses rows across multiple Geant4 simulations).

What changes

photon_afterpulse loops over ~16 PMT element types per chunk. In the
existing code, the inner work for each element is a chain of NumPy
ops — in-place divide, fancy-indexed gather of prob_ap, np.where
selection, np.argmin of a broadcast |cdf - r| — each allocating
intermediate arrays of size N_photons or (N_sel, K_bins). On
high-yield sources the orchestration around those NumPy calls
dominates the plugin's wall time.

The new design splits the inner work into two compiled kernels:

• _ap_select_kernel — one tight loop over N rows that scales
  rU0 in place by the per-element modifier (halving the DPE rows),
  compares each row against its per-channel probability ceiling, and
  emits the indices that pass the selection plus the running max of
  the ceiling (for the existing > 0.5 warning check). Replaces four
  separate numpy passes with a single integer-emitting loop.
• _ap_kernel_nonuniform — for the non-Uniform branch, two
  register-resident O(K) scans per selected photon (delay-time CDF
  and amplitude CDF). Replaces the broadcast-+-np.argmin pattern
  that materialised an (N_sel, K_bins) intermediate per call.

The amplitude CDF is sometimes 1D and sometimes 2D depending on the
element; numba does not accept a single parameter whose rank varies
per call, so both shapes are always passed and the kernel branches
on a flag.

All rng.random and rng.uniform draws stay in Python, in the same
order with the same argument shapes and the same conditional skip on
empty per-element selections — so the PCG64 sequence is preserved
exactly. The continue on n_sel == 0 must stay before the second
draw or the random sequence diverges; a comment in the source flags
this.

Bit-identity verification

pmt_afterpulses output was compared field-by-field against the
unmodified plugin on K40, Co60, Th232, Ra226, and Pb212
Geant4 simulations — ~30 M total rows, every field
(channel, dpe, photon_gain, cluster_id, photon_type,
time, endtime) byte-identical.

Performance

Per-source pmt_afterpulses target wall, before → after:

source	vanilla (s)	this PR (s)	speedup
K40	49.7	30.3	1.64×
Co60	650.1	380.3	1.71×
Th232	28.7	17.1	1.68×
Ra226	35.9	21.6	1.66×
Pb212	1046.4	771.2	1.36×

Notes

• Single-file diff: only
  fuse/plugins/pmt_and_daq/pmt_afterpulses.py is touched.
• First-call JIT compile of the two new kernels is one-time per
  process (cache=True persists across invocations).