fix(asr): stop evaluate_eou returning negative false_negatives and missing

[Kayvan-Zahiri]

evaluate_eou can return negative false_negatives and missing, which are counts.

When a prediction cuts a reference short, the reference index goes into earlycut_ids
and r_idx stays put. If a later prediction starts after that reference ends, the
while loop advances r_idx past it, so the trailing "unmatched references" block
never counts it. The unconditional -= len(earlycut_ids) still subtracts it, and both
counters go below zero.

Two references, one early cutoff followed by an exact hit:

reference  = [{"start_time": 0.0, "end_time": 2.0}, {"start_time": 5.0, "end_time": 7.0}]
prediction = [{"start_time": 0.0, "end_time": 1.0, "eou_prob": 0.9},
              {"start_time": 5.5, "end_time": 7.0, "eou_prob": 0.9}]
evaluate_eou(prediction=prediction, reference=reference, threshold=None, collar=0.1)
# false_negatives = -1, missing = -1

The fix discounts only the early cutoffs the trailing block actually counted, so a
trailing early cutoff is still discounted as before. Two regression tests added; the
first fails on main, the second guards the existing behavior.

Existing tests/collections/asr/test_asr_eou.py passes before and after.

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.