Add support for HuggingFace's TensorFlow models

[valeriosofi]

New Feature:

This PR implements support for HuggingFace's TensorFlow models, solving #125.
It also adds a notebook that shows a use case with BERT.

Example of usage:

from speedster import optimize_model
from transformers import TFAutoModel, AutoTokenizer

# Make sure to always load a matching tokenizer and model!
tokenizer = AutoTokenizer.from_pretrained("bert-base-cased")
model = TFAutoModel.from_pretrained("bert-base-cased")

# Let's load some data and tokenize it
test_string = "This is a sentence!"
tokenized_inputs = tokenizer(test_string, return_tensors="np")

dynamic_info = {
    "inputs": [
        {0: "batch", 1: "num_tokens"},
        {0: "batch", 1: "num_tokens"},
        {0: "batch", 1: "num_tokens"},
    ],
    "outputs": [{0: "batch", 1: "num_tokens"}, {0: "batch"}],
}

# Run Speedster optimization
optimized_model = optimize_model(
  model, input_data=[tokenized_inputs], dynamic_info=dynamic_info
)

Minor changes:

• Added a warning when dynamic_info is not provided with HuggingFace models
• The final report on the optimized model now contains also the name of the compiler used
• Fix a tvm issue when using the PyTorch implementation with gpu
• Added dynamic shape support for optimized models converted from TensorFlow to ONNX
• Fixed an issue that caused a wrong output shape for the models optimized with the neural compressor
• Enabled openVINO compiler when using python 3.10

[nebuly-ai]

LGTM!

[diegofiori]

@valeriosofi why was the azure test cancelled?

[valeriosofi]

@valeriosofi why was the azure test cancelled?

Because there was a bug, so there was no reason to perform the tests, they would have failed

[diegofiori]

Closes #125 and #103