Type the two error conventions crossing the hybrid-pipeline boundary

[LegrandNico]

In pyhgf's "hybrid" / mixed pipelines, you can chain together two kinds of building blocks:

• Standard deep-learning parts (like a GELU activation or layer-norm) that pass errors around the way PyTorch/TensorFlow/JAX do, the "gradient" convention, where a positive error means "the output was too high."
• PyHGF learning parts that use the framework's native belief-updating math, the "prediction error" convention, where a positive error means "the observation was higher than the prediction."
  These two conventions are exact negatives of each other. If you accidentally mix them up the code runs fine, but training silently diverges. That's a nasty class of bug: invisible, and only detectable by noticing the model won't learn.

This PR makes this sign convention explicit and self-documenting so those bugs are caught or avoided, without changing any runtime behavior:

DescentError (gradient-like: positive = output too high) and ObservedMinusPredicted (PyHGF's prediction error: positive = observation exceeded prediction) are now distinct annotated types, with validate_error_convention as a runtime sanity check. The hybrid-pipeline docstrings spell out where the single descent ↔ observed-minus-predicted conversion happens (the executor's backward pass for DeepNetworkAdapter), and EquinoxAdapter.backward_fn's signature declares the convention it must respect. A missing negation at this boundary trains silently in the wrong direction; the types make that mistake visible.

[codecov]

Codecov Report

✅ All modified and coverable lines are covered by tests.

Files with missing lines	Coverage Δ
pyhgf/model/__init__.py	87.50% <100.00%> (+0.83%)	⬆️
pyhgf/model/error_types.py	100.00% <100.00%> (ø)
pyhgf/model/hybrid.py	86.90% <100.00%> (+0.15%)	⬆️

🚀 New features to boost your workflow:

• ❄️ Test Analytics: Detect flaky tests, report on failures, and find test suite problems.