refactor: remove coremltools as required dependency and introduce numba kmeans1d

[guru-desh]

Description

This PR removes coremltools as a required dependency for coreai-opt.

The only part of coremltools that was being used was kmeans1d. We have re-implemented kmeans1d in coreai-opt using numba to keep coreai-opt platform-agnostic.

numba is a JIT compiler for Python that works by adding decorators to Python functions. This is different compared to coremltools kmeans1d that relied on AOT compilation of CPP code.

The alternative of using numba would be to lazily do AOT compilation in torch using torch.utils.cpp_extension. This hasn't been tested.

Performance Benchmarking

numba kmeans1d is around 1-3x faster than coremltools kmeans1d. This is mainly because coremltools kmeans1d doesn't compile with -O3. Once -O3 is used, we should see the performance be the same according to this numba forum post. AOT C++ is faster at small tensor sizes (10-20 elements) with low k as numba JIT compilation time is much higher than kmeans1d execution time.

I have attached a markdown file where I ran performance comparisons of numba and kmeans1d:
benchmark_results.md

Testing

1. We have equivalence tests to ensure that numba kmeans1d is equivalent to coremltools kmeans1d (tolerance is 1e-9)
2. kmeans1d tests are copied over from coremltools
3. These tests pass in CI

[guru-desh]

Currently, if k > n, the behavior is to get the minimum of the two. This is what the behavior was in coremltools's kmeans1d. We can choose to change it if wanted. I didn't change it because I wanted this PR to be as close to a pure refactor as possible