Add Numba JIT acceleration for KS computation (~2-3× end-to-end speedup)

[TorbenOestergaard]

Description:

This PR adds optional Numba JIT compilation to the inner loops of SAtom, significantly reducing computation time for large datasets.

Changes

_core.py

• Added three @‌njit(cache=True) functions:
  • _ks_row_mode01 — fused KS distance for SScompare 0 and 1
  • _ks_row_mode2 — fused KS distance for SScompare 2
  • _build_mask — behavioural mask construction with zero temporary array allocations
• Replaced YtempSorted (list of augmented arrays with 1-based MATLAB indices) with YsortedIdx — a clean (nOut, N) array of 0-based sort indices, used by both code paths
• Numba import is optional: if not installed, falls back to pure NumPy with a warning
• Added use_numba: bool = True parameter to SAtom() for easy A/B benchmarking

pyproject.toml

• Removed pytest, pandas, openpyxl from runtime dependencies (not required by satom)
• Added [fast] optional extra: pip install satom[fast] installs Numba
• Moved test/dev tools to [dev] extras

Performance

Tested on the Ishigami-Homma function (3 inputs, 1 output, N=500) and the Viborg dataset (10 inputs, 3 outputs, N=5000):

Benchmark NumPy Numba Speedup
Isolated KS row (Ishigami). ~500 µs. ~31 µs. 16×
End-to-end SAtom (Viborg, J=1000). ~0.9 s. ~0.3–0.5 s. 2–3×

The first call incurs a one-time JIT compilation cost (~1–2 s); subsequent calls use the cached compilation.

Backward compatibility

• All 19 existing tests pass unchanged
• use_numba=False reproduces the original NumPy behavior exactly
• Without Numba installed, the package works as before (with a warning)