Add option to use bf16 in PT sdp (#5)

[astachowiczhabana]

The new option --sdp_on_bf16 allows pyTorch to use reduced precision in sdp in the math backend. The change affects the following examples:

• contrastive-image-text/run_bridgetower.py
• question-answering/run_qa.py
• speech-recognition/run_speech_recognition_ctc.py
• stable-diffusion//text_to_image_generation.py
• text-classification/run_glue.py

[astachowiczhabana]

Hi @libinta this commit is also required with next OH release

[jiminha]

@astachowiczhabana which model and configuration are you running with this flag enabled? We'd like to do the same for the OH CI/PYTEST.

[hsubramony]

@astachowiczhabana dont we need to update README for all the models with this option ?

[ugolowic]

@jiminha

Inference - 1 card:

• question-answering/run_qa.py bert-large-uncased-whole-word-masking bs 8
• text-classification/run_glue.py bert-large-uncased-whole-word-masking bs 8
• stable-diffusion/text_to_image_generation.py stabilityai/stable-diffusion-2-base bs 4
• speech-recognition/run_speech_recognition_ctc.pywav2vec2-librispeech-clean-100h-demo-dist bs 128
• contrastive-image-text/run_bridgetower.py BridgeTower/bridgetower-large-itm-mlm-itc bs 16

Training - 8 cards:

• audio-classification/run_audio_classification.py facebook/wav2vec2-base bs 16
• image-classification/run_image_classification.py google/vit-base-patch16-224-in21k bs 64