fix: analysis log_likelihood_function preserves the cause behind FitException

[Jammy2211]

Summary

The numpy-path log_likelihood_function in all three lens analyses wrapped fit
failures in except Exception as e: raise af.exc.FitException and discarded
e. The FitException is correct (PR#607 path-parity: a numpy non-PD
LinAlgError → FitException → the sampler resamples, matching the JAX
NaN-resample), but the bare re-raise masked the true error — the traceback
showed only FitException at analysis.py, and a genuine code bug was
indistinguishable from a pathological-model resample. Surfaced by
autolens_workspace#308/#309 (interferometer-Delaunay marker re-diagnosis).

Fix: a shared raise_fit_exception helper chains the cause (raise ... from e)
and honours an opt-in PYAUTO_RAISE_ANALYSIS_EXCEPTIONS=1 to re-raise the
original exception unwrapped for debugging. Default behaviour is unchanged —
production searches still resample on FitException.

API Changes

Changed Behaviour

• AnalysisImaging / AnalysisInterferometer / AnalysisPoint
log_likelihood_function (numpy path): a fit failure is still re-raised as
  af.exc.FitException, but now with the original exception preserved as
  __cause__ (raise ... from e). No change to when/whether FitException is
  raised; searches resample exactly as before.

Added

• Env var PYAUTO_RAISE_ANALYSIS_EXCEPTIONS=1 (default off / "0"): re-raises
  the original exception unwrapped instead of wrapping it — a debug aid for
  surfacing masked failures (e.g. under PYAUTO_TEST_MODE, where a single
  evaluation is not absorbed by a sampler).
• autolens.analysis.exceptions.raise_fit_exception(exception) — internal
  helper implementing the above (not part of the user-facing modelling API).

Removed / Renamed / Changed Signature

• None.

Migration

• None required. Existing code and configs are unaffected; the only observable
  difference is a richer traceback (__cause__) on an already-raised
  FitException.

Test Plan

• New test_autolens/analysis/test_exceptions.py (3 tests): wraps + preserves
  __cause__; env flag re-raises original; "0" still wraps (no truthy-string trap).
• Existing analysis/model tests pass (16) — no regression.

Refs #638.

Generated by the PyAutoLabs agent workflow.