Add ONNXRuntime support

[tanaysoni]

This PR adds the conversion of PyTorch BERT models to ONNX format for inference with ONNXRuntime.

Speed benchmarks: https://docs.google.com/spreadsheets/d/1ak9Cxj1zcNBDtjf7qn2j_ydKDDzpBgWiyJ7cO-7BPvA/edit#gid=0

Usage:

from haystack.reader.farm import FARMReader
from pathlib import Path

FARMReader.convert_to_onnx(
    model_name_or_path="deepset/bert-base-cased-squad2", optimize_for="gpu_tensor_core"
)  # exports the model to ./onnx-export dir
FARMReader(model_name_or_path=Path("onnx-export"))  # load the ONNX model

[tholor]

Looking good once the merge conflict is resolved

[datalee]

support dpr model？

[Timoeller]

Not yet, for this we would need the convert_to_onnx method defined in the AdaptiveModel implemented in the BiadaptiveModel

[akmalmasud96]

Hi,
is there support available for the EmbeddingRetrieval in onnx?