Require evaluation mode for attribution patching

[emerardd]

Description

Require attribution_patch() to run with the model and every submodule in evaluation mode, preventing separate clean and corrupt forwards from mixing independent dropout realizations into false causal effects.

This extracts the all-submodule state validation already used by JacobianLens.fit() into a shared private helper. The helper also inspects a Bridge's deliberately unregistered original_model, reports the offending module paths, and rejects invalid state without mutating caller-owned training flags. Input-shape validation still runs first, while the state guard runs before either attribution forward.

Regression coverage includes root training mode, an eval root with only a nested dropout in training mode, hidden original_model state, preservation of training flags, and exact-zero attribution for identical inputs in evaluation mode.

Fixes #1787

Type of change

• Bug fix (non-breaking change which fixes an issue)

Checklist:

• I have commented my code, particularly in hard-to-understand areas
• I have made corresponding changes to the documentation
• My changes generate no new warnings
• I have added tests that prove my fix is effective or that my feature works
• New and existing unit tests pass locally with my changes; only the affected unit-test surface was run (142 passed)
• I have not rewritten tests relating to key interfaces which would affect backward compatibility

Local validation

pytest tests/unit/tools/test_attribution_patching.py tests/unit/tools/test_jacobian_lens.py -q
142 passed, 3 warnings in 27.65s

pycln --check --all: passed
isort --check-only: passed
black --check: passed
mypy: Success: no issues found in 399 source files
git diff --check: passed

[jlarson4]

Looks good, thanks @emerardd