fix(asr): stop silently dropping utterances and remove scoring hot spots

[liutaocode]

Problem

Three silent-failure paths in the ASR scoring pipeline understate error rates or report perfect scores on broken inputs:

1. Empty hypothesis text is dropped. aispeech_norm keeps only rows that split into two tab-separated fields, so an empty recognition result (key\t) is discarded, and compute_wer then skips the matching reference row (fid not in rec_set). The worst utterances — where the model produced nothing — are excluded from CER/WER/MER entirely. On a benchmark set where ~half the hypotheses are legitimately empty, corpus CER is understated by up to ~50%.
2. Missing hypothesis utterances are skipped silently, with no utterance counts anywhere in the report, so a partial or key-mismatched hypothesis file passes as a clean run.
3. Malformed input scores 0.0. A space-separated (Kaldi-style) file parses to zero rows, and all == 0 maps to a perfect 0.0 error rate with no warning.

Changes

• normalization/aispeech_norm: keep empty-text rows (scored as deletions downstream), report row_stats (rows / empty / dropped-malformed) in the node trace, and raise when a file yields no parseable <key>\t<text> rows.
• scoring/wenet_* wrapper: score reference utterances missing from the hypothesis as empty hypotheses (pure deletions), expose utterance coverage (num_ref_utts, num_hyp_utts, num_hyp_missing_utts, num_hyp_extra_utts, key samples) in the scoring result, and refuse to report 0.0 when zero reference tokens were covered. Code-switch zh/en side scores stay lenient (monolingual subsets legitimately cover zero tokens).
• Performance, bit-identical scores: reset only the DP submatrix each utterance uses in the vendored compute_wer (one long utterance previously made every following utterance pay O(longest²); 17x on a synthetic corpus, ~10x end-to-end on a real 1.4k-utterance code-switch run), and preload normalization map tables once per node call instead of re-globbing/re-reading per line or token (~40x per-call overhead).
• Fix a latent NameError in asr_num2words' fallback map loading (load_and_sort_map referenced but never defined).

Tests

Nine new regression cases (tests/test_asr_scoring_coverage.py) covering empty/missing hypotheses, coverage reporting, zero-token rejection, order-independent totals, and monolingual code-switch side scores. Legacy parity tests now subset-compare: scoring results may add coverage fields, but every legacy field must still match exactly.