Fix CTC model class resolution by passing model_type to TasksManager

[codykk]

Pass model_type to Optimum's TasksManager.get_model_class_for_task in resolve_task() Stage 2, fixing model class resolution for all CTC-based ASR architectures (Wav2Vec2ForCTC, HubertForCTC, WavLMForCTC, etc.). One-line generalized fix — no per-model hardcoding. Add regression recipe for jonatasgrosman/wav2vec2-large-xlsr-53-russian (cpu/fp32, cpu/fp16).

Effort: L1-light | Goal: L1 (perf) | Outcome: L1

Model metadata

• What it does: Wav2Vec2-Large-XLSR-53 fine-tuned on Russian Common Voice for CTC-based automatic speech recognition. Converts 16kHz raw audio waveforms to Russian text transcriptions.
• Primary user stories: Russian speech-to-text transcription for voice applications
• Supported tasks: automatic-speech-recognition (CTC decoding)
• Architecture:
  Wav2Vec2ForCTC
  ├── Wav2Vec2FeatureEncoder (7 CNN layers: raw audio → features)
  ├── Wav2Vec2FeatureProjection (LayerNorm + Linear → hidden_size=1024)
  ├── Wav2Vec2Encoder (24× transformer layers, 16 heads)
  └── Linear CTC head (1024 → 39 Cyrillic vocab)

Validation and support evidence

Baseline (main): Build FAIL — Unrecognized configuration class Wav2Vec2Config for this kind of AutoModel: AutoModelForSpeechSeq2Seq

Root cause: TasksManager.get_model_class_for_task("automatic-speech-recognition") without model_type defaults to AutoModelForSpeechSeq2Seq; CTC models need AutoModelForCTC.

Goal ceiling: L1 (perf) — L2/L3 CLI-BLOCKED (automatic-speech-recognition not in winml eval supported tasks)

Outcome: L1 — generalized code fix + regression recipes

EP	Device	Precision	Build	Perf
cpu	cpu	fp32	✅ PASS	125.6ms avg, 7.96 samples/sec, +1268MB RAM
cpu	cpu	fp16	✅ PASS	140.5ms avg, 7.12 samples/sec, +1276MB RAM

Delta: resolution.py line 647 — added model_type=model_type_norm or None. When model_type_norm is available (from config.model_type), Optimum looks up the correct class per architecture family instead of defaulting to the first registered class. Empty string falls back to None (original behavior preserved).

Reproduce:

winml build -m jonatasgrosman/wav2vec2-large-xlsr-53-russian -o output/ --ep cpu --device cpu
winml perf -m output/model.onnx --ep cpu --device cpu

[xieofxie]

Please pass the raw model_type to TasksManager here rather than model_type_norm. The normalization is needed for WinML's internal mapping keys, but Optimum registers some model types with underscores. For example, speech_to_text is a supported Transformers model type; converting it to speech-to-text prevents Optimum from matching AutoModelForSpeechSeq2Seq, so explicit ASR resolution raises instead of resolving the class. The Stage 2 call added by this PR has the same issue. Please keep model_type_norm for _get_custom_model_class, pass model_type to both TasksManager calls, and add an underscore-containing regression case.