Add Paraformer-zh ASR model support with OpenVINO inference and INT8 quantization

[padatta]

This PR adds comprehensive support for Paraformer-zh (FunASR) automatic speech recognition models in optimum-intel, including:

1. OpenVINO inference model (OVParaformerForSpeechSeq2Seq) following the established patterns from OVModelForTextToSpeechSeq2Seq
2. INT8 quantization support with NNCF weight compression during export
3. GPU optimization with dtype fixes and dynamic shape handling for Intel ARC/iGPU

What's Changed

1. Paraformer Inference Model (modeling_speech2text.py)

• New OVParaformerForSpeechSeq2Seq class for running Paraformer models with OpenVINO
• Supports both single-model and component-based architectures (encoder, predictor, decoder)
• Compatible with FP32, FP16, and INT8 quantized models

2. INT8 Quantization Support

• Integrated INT8_SYM weight compression using NNCF during export
• Export command properly handles Paraformer quantization workflow

3. GPU Compatibility Fixes

• Fixed dtype issues: use int32 for positions/ranges, int64 for large indices
• Implemented dynamic mask creation using shape operations (ONNX/OpenVINO compatible)
• Added max_label_len clamping (4096) to prevent memory issues

4. Comprehensive Test Suite

Added tests/openvino/test_paraformer.py with 10 test cases:

• Model loading from pretrained path
• Basic forward pass inference
• Batch inference with variable lengths
• NumPy array input support
• generate() API compatibility
• CPU ↔ GPU device switching
• Model serialization (save/load)
• Decoding logic without token_num
• Model properties validation
• Output dataclass structure

Usage Example

from optimum.intel.openvino import OVParaformerForSpeechSeq2Seq
import torch

# Load INT8 model on GPU
model = OVParaformerForSpeechSeq2Seq.from_pretrained(
    "path/to/paraformer-zh-int8/ov_models",
    device="GPU"
)

# Prepare input (extracted features from audio)
speech = torch.randn(1, 100, 560)  # [batch, time_frames, features]
speech_lengths = torch.tensor([100], dtype=torch.int32)

# Run inference
output = model(speech, speech_lengths)

# Access results
print(output.logits)      # [batch, seq_len, vocab_size]
print(output.token_ids)   # [batch, seq_len] - decoded tokens
print(output.token_num)   # [batch] - number of tokens per sequence

Export to OpenVINO with INT8 Quantization

optimum-cli export openvino \
  --model damo/speech_paraformer-large-vad-punc_asr_nat-zh-cn-16k-common-vocab8404-pytorch \
  --task automatic-speech-recognition \
  --weight-format int8 \
  paraformer-zh-int8

Testing

All 10 test cases pass:

export PARAFORMER_TEST_MODEL=/path/to/paraformer-zh/ov_models
python -m unittest tests.openvino.test_paraformer -v

test_batch_inference ... ok
test_decode_without_token_num ... ok
test_device_switching ... ok
test_generate_api ... ok
test_load_model_from_pretrained ... ok
test_model_inference ... ok
test_model_output_dataclass ... ok
test_model_properties ... ok
test_numpy_input ... ok
test_save_and_load ... ok

----------------------------------------------------------------------
Ran 10 tests in 13.071s

OK

Checklist

• Implementation follows existing optimum-intel patterns
• Comprehensive test suite added (10 test cases)
• All tests passing
• GPU and CPU support verified
• INT8 quantization validated
• Performance benchmarks included
• Code follows style guidelines
• Documentation inline with code

Note: This PR builds upon the existing Paraformer export support (commit f79a331) and adds the inference runtime capabilities to make the exported models usable within the optimum-intel framework.