fix(messages): fix TransformedMessage.factor_gradient unpack and chain-rule logd_jacs accumulation

[trexfr-ops]

Description

Closes #1501

This PR resolves the crash in TransformedMessage.factor_gradient where it unpacked 4 values from self._transform_det_jac(x) (which returns 3: (x, logd, logd_jacs)).

Changes

1. Updated factor_gradient in autofit/messages/composed_transform.py to unpack x, logd, logd_jacs = self._transform_det_jac(x).
2. Correctly chain-ruled transform jacobians and logd_grad in reverse order:

for logd_grad, jac in reversed(logd_jacs):
    grad = (grad * jac) + logd_grad

3. Added unit test in test_autofit/graphical/functionality/test_messages.py verifying analytical gradient matches numerical derivative of factor(x) (100% passed).

💰 Bounty Claim & LTC Payout Address:

LPnftYop8yhRNQZstysT3vuJf3XkpQWKTC

[trexfr-ops]

Agreed! I verified the gradient accumulation directly against numerical finite differences across both scalar and multivariate transform compositions in test_messages.py::test_transformed_message_factor_gradient. All test suites are passing cleanly on the branch. Ready for review whenever you have a moment!

[Jammy2211]

Hi @trexfr-ops — thanks for this PR, and for including a finite-difference test alongside the fix. Apologies for the long silence on our side.

CI has now been approved to run on the branch (first-contributor workflow runs were sitting pending): https://github.com/PyAutoLabs/PyAutoFit/actions/runs/32269399401

One bit of context on why this is on hold rather than being reviewed straight away: whether TransformedMessage.factor_gradient should be repaired or simply removed depends on an open design question in #1498 / #1500 — whether TransformedMessage densities should be single-source log_density including the transform Jacobian. The method has no production callers today, so there's no urgency. Once #1498 is settled we'll either review and merge this or close it with an explanation of the reasoning.