Add python example w/ cffi-generated bindings

[ochafik]

I've played around with Python bindings generation and thought it was worth adding as an example. This PR features:

• Access to full C API (incl. CUDA, MPI, OpenCL, Metal, alloc... and any local API changes)
• Instant regeneration with python regenerate.py (a cffi wrapper; uses llama.cpp headers by default, see README.md for options)
• Lightweight utils to do copies between tensors (ggml & numpy.ndarray alike) with automatic (de/re)quantization, or view a tensor as a numpy ndarray
• Full stubs with preservation of original signatures as docstrings (for neat autocomplete in IDEs)

You can play with it directly with this Colab

Some notes:

• I've committed the generated bindings (ggml/cffi.py) and stubs (ggml/__init__.pyi) mostly for shows (to ease up the review) but we probably won't want to keep them in the repo (lemme know what you think, happy to add a cmake command to generate them during the build)
• There's already bindings w/ high-quality docs (https://github.com/abetlen/ggml-python) but I wanted something that's easy to autogenerate / keep up-to-date with local changes, and I thought having it in the repo's examples made sense.
• I've tried cffi's native extension generation (see complex branch) and decided not to go for it as it's quite a bit more fiddly than this "little" example, with unclear performance benefits.

Apologies if my Python style is questionable, I'm a bit new 😅

[ggerganov]

Thanks for the contribution!

Looks interesting. I'm not a big Python user, so cannot tell how useful this is, but it looks well done. As long as it does not interfere with the C library we can add it to the repo.

Feel free to merge it

[ochafik]

@ggerganov thanks a lot, will merge now (pushed some cosmetic changes + regenerated bindings, so much new gguf stuff 🤗)

re/ usefulness, I'm working on a minimalist Python hybrid of llama2.c + llama.cpp, which I hope could be a easy platform to experiment with various modifications (e.g. on-the-fly matrix decomposition / pruning at loading time w/ numpy)

Debugging GGML_ASSERT failures in Python can be fiddly but I'll cleanup a util that catches SIGABRT and enters an interpreter for easy debugging.