What does this PR do?

This PR refactors the minDCF evaluation metric to resolve Out-Of-Memory (OOM) errors during the evaluation of large datasets.

Due to the previous memory footprint, the original function crashed when attempting to compute minDCF for large trial lists, such as VoxCeleb1-H (Vox-H) and VoxCeleb1-E (Vox-E).

Complexity Improvements:
Where N is the total number of scores and T is the number of unique thresholds (typically T ≤ N_pos + N_neg):

Old Implementation:

• Space: O(N * T)
• Time: O(N * T)

New Implementation:

• Space: O(N)
• Time: O(N log N)

Algorithmic Details:

1. Memory Optimization: Swapped the O(N * T) tensor expansion blocks for 1D torch.searchsorted operations to count False Acceptance and False Rejection rates.
2. Removed Midpoints: The calculation of intermediate thresholds ((thresholds[0:-1] + thresholds[1:]) / 2) was removed. Because p_miss and p_fa are step functions that only change state at observed scores, evaluating midpoints is mathematically redundant.

Testing:

• Empirical Parity: Evaluated on the VoxCeleb1-O (Vox-O) test set. The optimized function yields the exact same minDCF result as the original implementation.
• Passes pytest tests/unittests/test_metrics.py.
• Passes pytest --doctest-modules speechbrain/utils/metric_stats.py.

Fixes #<issue_number>

Breaking Changes:
No breaking changes. The function signature and return types remain identical.

PR review