Add spice -> kernel pipeline (spice_kernel) for the --shape fast constants

[WillMatthews]

Second of two stacked PRs. Stacked on #30 (archive) — review/merge #30 first; the base will retarget to master (via fast-output-shaping) as the stack lands.

This PR (pipeline): src/spice_kernel.cpp — a C++ tool that closes the loop from the circuit to the shaper. It:

1. runs spice/jseries_fastout.cir via ngspice and extracts the single-photon fast-output pulse;
2. derives the two --shape fast time constants — tauLoad = R_Lfast·N·C_f (read from the deck's .param lines, so it can't drift from the circuit) and tauRecovery (fitted on the fast tail);
3. validates the analytic kernel shape_fast() implements against the SPICE pulse;
4. writes a ready-to-run sipm_fast.json (via the existing save_params_json) whose tauLoad is circuit-calibrated.

Built via make spice_kernel (excluded from the default all link, like server.cpp):

$ make spice_kernel && ./build/apps/spice_kernel
=== spice -> --shape fast kernel ===
  circuit: N=14410, C_f=3.47 fF, R_Lfast=50 ohm, OV=2.5 V
  tauLoad     = R_Lfast*N*C_f =   2.500 ns   (-> params 'tauLoad')
  tauRecovery = loaded fit    =  24.539 ns   (fast-tail recovery)
  kappa       = C_f/(C_d+C_q) = 5.4219e-02
  FWHM: spice 1.93 ns vs analytic kernel 1.51 ns
  peak-normalised shape RMS error (kernel vs spice, <40 ns): 0.088
  [PASS] shape RMS 0.088 < 0.10
  wrote sipm_fast.json

Then simspad -p sipm_fast.json -i light.npy -o resp.npy --shape fast gives a fast-output trace calibrated to the circuit rather than to a hand-picked constant (verified to round-trip through simspad). Needs ngspice on PATH; the generated params file is gitignored.

C++ only — no Python added to the package.

Closes #29.