Metalcompute

What's Changed
* Update to python 3.12 and latest clang

- Resolve build failure with clang 14 (23)
- Resolve potential memory leakage (22)
- Resolve issue when creating buffers with array of length 1 (21)
- Update wheel build to cover all python versions

- Ensures MTLBuffer objects are marked as empty before releasing reference to ensure mapped memory is reclaimed (19)

- Resolves an issue with reference counting of temporary buffers created when passing python buffers to kernel functions directly
- Adds video encode example to mandelbrot notebook

- Resolves issue 13 - contiguous numpy buffers can now be passed directly as kernel call inputs or buffer initializers

- New object based interface
- README and examples all updated to new interface
- Old interface still available but will be removed in later release
- Can run kernels with 1 or more buffers, input/output defined by kernel program
- Zero copy buffers supporting python buffer interface, allowing direct use with numpy using np.frombuffer
- Can pipeline kernel requests to keep device busy
- metalcompute-measure-flops script example now renamed to -measure, includes TFLOPS & data transfer bandwidth