perf: profile + speed up AggregateImages/AggregateFITS (png/fits workflow aggregators)

[Jammy2211]

Overview

Final leg of the aggregator profiling arc (#1375 A–D + awt#171 merged): only AggregateCSV of the three catalogue workflow aggregators was profiled and sped up (−25% in #1376). AggregateImages (png_make) and AggregateFITS (fits_make) were never measured — and code reading already shows AggregateFITS._hdus re-opens the same fits file once per requested HDU via result.value(...) (an uncached fits.open per HDU) and never closes handles.

Plan

• Harness (autofit_workspace_test): give mock templates realistic payloads — a panelled subplot_fit.png and a multi-HDU fit.fits — and add aggregate_images (output_to_folder, the png_make pattern) + aggregate_fits (extract_fits + output_to_folder, the fits_make pattern) stages to the profiling grid.
• Baseline, then low-hanging fruit in autofit/aggregator/summary/aggregate_images.py / aggregate_fits.py (candidates: one fits.open per source file per result; close handles; avoid repeated result.value lookups). Behaviour-preserving; before/after in the PR.
• Out of scope: faster single-file deserialization (known deeper follow-up).

Worktree: ~/Code/PyAutoLabs-wt/aggregate-images-fits/ · Branch: feature/aggregate-images-fits · Repos: PyAutoFit (primary), autofit_workspace_test — both clean on main, unclaimed.

[Jammy2211]

Shipped and merged (2026-07-17): #1386 + PyAutoLabs/autofit_workspace_test#51.

What the png/fits profiling found:

• AggregateImages (png_make): ~1.5 ms/result, already caches one open per subplot type per result — profiled clean, no fruit taken.
• AggregateFITS (fits_make): leaked one file handle per requested HDU per result (200 fds measured at 100 results × 2 HDUs). At the default ulimit of 1024, extract_fits/output_to_folder crashes with "Too many open files" around ~500 results — the 3000-lens Euclid catalogue would have hit this. extract_csv leaked one handle per result the same way.
• Fixed: one open per source file per result, data copied out of the memmap, deterministic close. 0 leaked fds at 100 results, identical output, 19% faster at 2 HDUs (interleaved A/B; grows with HDU count — the workflow guide uses 4).

This closes the aggregator arc's last uncovered leg: all three catalogue workflows (csv −25% in #1376, png profiled-clean, fits leak-fixed +19%) are now measured and harnessed. Closing.