profiling: standardise on NUFFT + sparse operator across interferometer/datacube

[Jammy2211]

Summary

• All interferometer/datacube profiling scripts now use TransformerNUFFT + apply_sparse_operator(use_jax=True) — the production-relevant likelihood path on A100.
• Added per-step sparse-mode breakdown for datacube (replaces the previous aggregated "sparse pipeline" step with individually-profiled D, F, reconstruction, log-evidence steps using the W̃ FFT kernel).
• Bumped Hilbert vertex count from 500/1000 to 1500 across delaunay scripts (matches imaging fiducial).

Context

apply_sparse_operator calls self.transformer.image_from(...) to compute the dirty image at setup time. With DFT this is O(N_pix × N_vis) — prohibitive at ALMA (1M vis × 800² pixels, hours on CPU). PyAutoArray#329 enabled NUFFT here; PyAutoArray#330 added chunking so 5M-vis ALMA_HIGH fits a 3 GB gather buffer per chunk via lax.scan.

Results (A100 80 GB, 34 channels, 1500 verts)

Config	Cube total	Curvature F	Shared L^T W̃ L savings est.
SMA fp64	1.74 s	0.21 s	0.20 s (11.7%)
ALMA fp64	2.26 s	0.74 s	0.72 s (31.8%)
ALMA HIGH fp64	4.82 s	3.15 s	3.06 s (63.5%)

Files changed

• likelihood_runtime/interferometer/pixelization.py — DFT → NUFFT via _build_transformer factory
• likelihood_breakdown/interferometer/delaunay.py — same, plus removed stale "pynufft is not JIT-friendly" comment
• likelihood_breakdown/datacube/delaunay.py — per-step sparse-mode breakdown (FFT W̃ curvature, sparse D, fast_chi²)
• likelihood_runtime/datacube/delaunay.py + likelihood_runtime/interferometer/delaunay.py — vertex bump to 1500

Test plan

• SMA fp64/mp + ALMA fp64/mp + ALMA_HIGH fp64/mp HPC breakdown jobs all passed rtol=1e-4 assertion against summed FitInterferometer.log_evidence
• All scripts pass AUTOLENS_PROFILING_SMOKE=1

🤖 Generated with Claude Code