Add Faster Transformer compiler for Bert

[cccntu]

WIP for #154

• [x] Code to install FasterTransformer
• [x] Code to optimize model using FasterTransformer

test this PR locally

WIP for #154

• [x] Code to install FasterTransformer
• [x] Code to optimize model using FasterTransformer

test this PR locally

• use the docker image (I'm not sure if the latest I have is actually the latest, it's id is 9e148a3a1d4f)

docker run --rm --gpus all -ti nebulydocker/nebullvm:latest

install FasterTransformer

git clone  <repo>
cd nebullvm
git checkout <branch>

# install speedster and nebullvm separately
# because speedster depends on nebullvm, it may try to install from pypi?

pip install -e apps/accelerate/speedster
pip install -e .
# probably need to checkout to another directory before running the below command
# so faster_transformer's library code is installed to the correct location
python -m nebullvm.installers.auto_installer  --f torch --compilers faster_transformer

• verify installation is successful

from nebullvm.operations.optimizations.compilers.utils import faster_transformer_is_available
faster_transformer_is_available()

• run the bert example

python notebooks/speedster/huggingface/faster_transformer_bert.py