The XY correlation cube snapped to center the moment the audio dropped
below the silence threshold, because process() hard-set value to 0.0 in
the silent branch. Now the silent branch eases the displayed value back
toward zero using the configured response time, so the cube holds its
last reading and then settles instead of jumping. The active path is
unchanged, so all existing response-time and peak tests still pass.
- add `display` (f64) that tracks the live correlation while active
- silent branch: display += alpha * (0 - display)
- update configurable_silence_threshold test for the eased decay
- add correlation_holds_then_decays_to_zero_on_silence test
