fix: assign sentence speakers by total overlap without double counting

[LauraGPT]

Problem

The shared CAM++ speaker assignment helper counts a newly leading segment's overlap twice and only accumulates segments for the currently leading speaker. For a 0-10 second sentence with speaker 0 at 0-4s and speaker 1 at 4-10s, the current real module returns speaker 0 despite speaker 1 covering 6 seconds.

Change

• Accumulate each positive, sentence-clipped segment overlap once per speaker; select the greatest total after all segments.
• Resolve equal totals deterministically using the first positively overlapping speaker. This explicitly defines ties rather than promising equivalence with every result of the old faulty algorithm.
• Preserve millisecond/second conversion, integer labels, no-overlap fallback 0, in-place sentence/list identity, and all other fields.
• Add 32 regression cases and wire them into both events and the test/format steps of the existing speaker-adapter workflow.

The normal caller supplies postprocessed diarization intervals. This does not add union/deduplication semantics for arbitrary overlapping or duplicate intervals.

Validation

• Real-module RED: the new 4s/6s test failed on exact main da12182; GREEN after the fix.
• 32 direct helper cases: nonconsecutive turns, all 24 order permutations, late-equalizing ties, boundary clipping, multiple sentences, in-place preservation, NumPy integer labels, empty input and boundary-only contact.
• 202 passed across tests/test_campplus_utils.py, test_server_app_openai_segments.py, test_realtime_ws_service.py, test_moss_transcribe_diarize_model.py, test_moss_transcribe_diarize_docs.py and test_qwen3_asr_vllm_offline_example.py.
• Python 3.12.3, existing torch 2.11.0+cu128 environment; no weights loaded, no GPU/model inference or new dependency installation.
• compileall funasr examples tests and git diff --check passed. Two existing invalid-escape SyntaxWarnings in the unrelated examples/industrial_data_pretraining/fun_asr_nano/tools/format5res.py remain.
• Independent review found no blocking issue; SSH-signed DCO commit.

Investigated while triaging #3727. This is an independently reproduced assignment defect, not a reproduction or resolution of that reporter's audio/UI symptoms. Keep that issue open for original audio/configuration evidence. No acoustic-accuracy, speaker-clustering or real-world streaming-quality claim.