fix: invalidate stale factor-graph subsample instances

[Jammy2211]

Release-blocking regression

Stage 3 rehearsal 31517075410 uses PyAutoFit b499d4367cfdf344af32d686246ded234f1ad74b (#1466), PyAutoGalaxy be61b8d0546c5d933ff9ef4bf9d445ba4595e03b (#568), TestPyPI 2026.8.11.1.dev71002, and the release profile (PYAUTO_TEST_MODE=1).

Valid factor-graph scripts still fail with child results missing per-analysis attributes such as .centre and .galaxies.

Root cause

start_resume_fit reads samples_summary.instance during test-mode validation, caching the global factor-graph instance in _instance. SamplesSummary.subsamples and Samples.subsamples shallow-copy the object and rebind it to a factor model, but retain that cached global instance. Child results therefore expose the global factor-index structure rather than the per-analysis model.

#1466's regression covered rejected final samples. Recovery creates a fresh summary, so it missed this valid-sample cache-warming path.

Plan

High level

• Invalidate cached instances whenever samples are rebound to a different model.
• Add direct cache-invalidation tests for summary and full samples.
• Add an end-to-end valid factor-graph regression in test mode.
• Preserve the PyAutoGalaxy parameter guards and the narrow FitException rejection contract.
• Run focused tests, the complete PyAutoFit suite, CI, and the complete Stage 3 rehearsal.

Detailed

• Update SamplesSummary.subsamples and Samples.subsamples in the samples interface implementation to clear _instance after copying and before/while rebinding the model/path metadata.
• Audit the adjacent with_paths/without_paths rebinding helpers for the same stale-cache invariant and cover/fix them if applicable.
• Add unit tests that first warm the global .instance, then derive a child samples object and assert its instance is reconstructed using the child model and is not the global cached object.
• Extend the existing fix: preserve guarded sample lifecycle #1466 factor-graph test area with a valid final-sample PYAUTO_TEST_MODE=1 case whose child results expose their expected per-analysis structure.
• Retain existing tests proving non-FitException programming errors propagate.

Suggested branch: fix/factor-graph-subsamples-cache.