fix(tests): put ecapa_tdnn models in eval mode before batch=1 forward passes

mlx>=0.32 tightened BatchNorm to reject training-mode calls with a
single value per channel. These unit tests construct EcapaTdnnBackbone/
ECAPA_TDNN directly (bypassing load_model(), which already calls
.eval()) and forward a batch=1 input while still in training mode,
so they now hit that guard. Production users are unaffected since
load_model() puts every loaded model in eval mode.

Co-Authored-By: Claude Sonnet 5 <noreply@anthropic.com>