fix: stop DER frame counts from saturating float16

[Atishyy27]

Closes #2042.

discrete_diarization_error_rate casts reference and hypothesis to np.half before counting frames, and np.sum accumulates in the array dtype. float16 tops out at 65504, so on a recording with more active frames than that the total saturates to inf and the error rate is reported as 0.0, roughly 18 minutes of single-speaker speech at the usual frame resolution, so most real recordings are affected, and a wrong output silently scores as a perfect one.

This accumulates the frame counts in float64. The np.half cast is left in place, so the comparison and memory behaviour are unchanged.

The same function divided by the total without checking it, so a reference with no speech gave inf or nan (reported in #1790, closed unfixed). The three divisions now go through a small helper that follows the convention already used by pyannote.metrics.detection: an empty reference scores 0 when nothing was detected either, and 1 otherwise.

Testing

Two tests in tests/test_metrics.py, both failing on develop:

• test_discrete_der_long_recording_does_not_overflow: 450000 frames (about two hours) with 1000 false-alarm frames. Returns 0.0 with total = inf before, 0.00222 with total = 450000.0 after.
• test_discrete_der_empty_reference: an all-zero reference returns nan / inf before, 0.0 / 1.0 after.

Also checked that a short recording is unaffected: 1000 frames with 100 false-alarm frames gives 0.1 before and after.