feat: CSV I/O for PointDataset for cluster-scale workflows

[Jammy2211]

Summary

Strong-lens cluster workflows can involve tens or hundreds of multiply-imaged
background sources. Editing that many per-source JSON files by hand is
unwieldy — one row per image grouped by name in a spreadsheet is a much
better hand-editing surface.

This PR adds a CSV I/O path for PointDataset alongside the existing JSON
path. JSON remains the canonical exact-round-trip format used by the
modeling scripts; CSV is the hand-editable cluster form. No new runtime
dependencies — csv stdlib only, no pandas.

Closes #446.

API Changes

• Added: PointDataset.to_csv(file_path), classmethod
  PointDataset.from_csv(file_path, name=None), and module-level
  autolens.output_to_csv(datasets, file_path) /
  autolens.list_from_csv(file_path) — CSV I/O, one row per observed image,
  grouped by name; optional flux / time_delay column groups dropped
  when blank for every row.
• Docstring fix: drop the stale PointDict bullet in
  autolens/point/dataset.py module docstring (that class no longer
  exists); document the JSON-vs-CSV surfaces instead.
• No removals, renames, or signature changes. Purely additive.

See full details below.

Test Plan

• test_autolens/point/test_dataset.py — round-trip tests for:
  • positions-only
  • positions + fluxes + fluxes_noise_map
  • positions + time_delays + time_delays_noise_map
  • all optional fields populated
  • heterogeneous list (one dataset with fluxes, one without) — exercises
    optional-column handling at the list level
  • multi-group from_csv requires explicit name=, single-group
    auto-picks
• Module-level manual round-trip of every case (pytest couldn't load in
  the sandbox: installed autogalaxy predates LensCalc, the conftest
  import fails before collection — not an issue with the new code).
• CI / maintainer-side pytest test_autolens/point/test_dataset.py

Full API Changes (for automation & release notes)

Added

• autolens.PointDataset.to_csv(file_path: str) -> None — write a single
  dataset as one CSV, one row per observed image.
• classmethod autolens.PointDataset.from_csv(file_path: str, name: Optional[str] = None) -> PointDataset
  — load a single dataset; name= required when the CSV has multiple
  groups, auto-picked when it has exactly one.
• autolens.output_to_csv(datasets: List[PointDataset], file_path: str) -> None
  — write a list of datasets to a single CSV; optional flux / time-delay
  column groups are included when any dataset carries those values and
  left blank for datasets that do not.
• autolens.list_from_csv(file_path: str) -> List[PointDataset] — load a
  list of datasets grouped by the name column in order of first
  appearance; raises ValueError on partially-populated optional columns
  within a group.

CSV schema (one row per image)

| name | y | x | positions_noise | flux | flux_noise | time_delay | time_delay_noise |

• name, y, x, positions_noise always written.
• flux, flux_noise added iff any dataset has fluxes.
• time_delay, time_delay_noise added iff any dataset has time_delays.

Changed Behaviour

• None for existing code paths.

Removed

• None.

Renamed

• None.

Migration

• No migration needed — JSON paths (output_to_json / from_json) are
  untouched and remain the canonical modeling input. CSV is opt-in for
  cluster-scale hand-editing workflows.

Non-goals (explicitly deferred to follow-up issues)

• FitPointDataset.to_csv residual / chi-squared export.
• Introducing a PointDatasetList class (list-level helpers stay as plain
  functions for now).
• Adding pandas as a dependency.
• Promoting CSV I/O to autoconf alongside fits / json (discussed with
  @Jammy2211 — reasonable future cleanup, but out of scope here).

https://claude.ai/code/session_01WUVZQGGLQhVKcprvUMf1Sa