fix(eeg): resolve NumPy trapezoid compatibility lazily

[aurascoper]

Scripts/eeg_spectral.py guarded its NumPy version bridge with:

_trapz = getattr(np, "trapezoid", getattr(np, "trapz"))

Python evaluates arguments before calling getattr, so the default runs
unconditionally. NumPy 2.0 removed trapz, so on precisely the version the
fallback exists for, importing the module raises:

AttributeError: module 'numpy' has no attribute 'trapz'

That takes down Scripts/consume-session.py and Tests/eval/test_session_consume.py
with it. The pinned calibration venv (numpy 1.26.4) hid the defect entirely.

Fix

Resolve lazily, and take the namespace as a parameter so both API shapes are
testable from a single interpreter:

def _resolve_trapezoid(namespace):
    trapezoid = getattr(namespace, "trapezoid", None)
    if trapezoid is not None:
        return trapezoid
    return namespace.trapz

Tests

Tests/eval/test_eeg_spectral_compat.py uses fake namespaces for both majors,
plus a TrapzIsFatal namespace whose trapz raises if touched — that one is
the actual regression guard and fails against the old expression.

Verification

environment	result
numpy 1.26.4 (pinned venv)	6/6 pass
numpy 2.4.6	6/6 pass
test_session_consume on numpy 2.4.6	crashed at import → 14/14 pass

Scope note

NeuralComposeEEG/src/neuralcompose_eeg/features.py is not affected — it
already guards with hasattr(np, "trapezoid") before reaching trapz, so its
fallback is only evaluated on numpy 1.x where it exists. I incorrectly reported
it as having the same exposure earlier; this PR touches one file plus its test.

Sequencing

Worth landing before the Python CI job: wiring Tests/eval into CI on a modern
NumPy would fail on this bug.