fix(loader): pass model_type to TasksManager for CTC ASR model class resolution

[codykk]

Fix incorrect model class resolution for CTC-based ASR models (Wav2Vec2, Hubert, Data2Vec-Audio). When automatic-speech-recognition maps to multiple candidate classes, Optimum defaults to AutoModelForSpeechSeq2Seq without model_type context, causing all *ForCTC models to fail. Passing model_type enables correct per-family resolution.

Add verified CPU recipes for jonatasgrosman/wav2vec2-large-xlsr-53-russian (fp32 + fp16).

Model metadata

• What it does: Large-scale cross-lingual speech representation model fine-tuned for Russian automatic speech recognition (CTC decoding)
• Architecture: Wav2Vec2ForCTC (wav2vec2) — CNN feature encoder → transformer encoder → CTC linear head
• Task: automatic-speech-recognition
• Vocab size: 39 (Russian characters + CTC blank)

Validation and support evidence

Baseline (main, 4daf0f1): FAIL — Unrecognized configuration class Wav2Vec2Config for AutoModelForSpeechSeq2Seq

Fix: Pass model_type=model_type_norm to TasksManager.get_model_class_for_task() (1-line change)

EP	Device	Precision	Build	Perf (avg latency)	Throughput
CPU	cpu	fp32	✅ PASS	87.10 ms	11.48 samples/sec
CPU	cpu	fp16	✅ PASS	—	9.49 samples/sec

Generalization confirmed: HubertForCTC, Data2VecAudioForCTC also resolve correctly. WhisperForConditionalGeneration (Seq2Seq) unaffected.

Reproduce:

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