Ggml exporter

[baskrahmer]

What does this PR do?

Export Hugging Face models to ggml format using a dedicated exporter. Right now I added the bloom and starcoder architectures, only for PyTorch and without quantization. In the current PR, adding a new architecture would look as follows:

• Adding a cpp file: contains code for inference and loading of the weights (example)
• Adding a Makefile target: used to create a main cpp binary from the cpp file (example)
• Creating a ggml config: used to create a ggml cpp binary from the Hugging Face weights. Specify parameters like vocab size, hidden dim etc. and also for model-specific processing (e.g. map HF layers to cpp layers using state dict; reshape specific layers, etc) -- see optimum/exporters/ggml/model_configs.py in this PR

Now I am a bit overwhelmed by the diversity between only starcoder and bloom already. This makes me question whether a solution like this would be too complex to support a wide range of models, hence I am opening it early as a draft PR to gather some feedback early before polishing any further.

To try it yourself, see these testcases tests/exporters/ggml/test_ggml_export.py and make sure the ggml.cpp submodule is loaded correctly (I think you need to explicitly pull it for the source code to appear).

Related issue: #903 (cc @fxmarty @NouamaneTazi @sidistic)

Before submitting

• Did you make sure to update the documentation with your changes?
• Did you write any new necessary tests?