Add mlx backend

[ifsheldon]

This resolves the issue in #295 (comment), adding a backend for MLX so that we can use einsum. Since MLX has very similar APIs to PyTorch and Numpy, the MLXBackend looks very similar to NumpyBackend and TorchBackend.

The tests passed on my Mac.

$ uv run einops/tests/run_tests.py mlx
running: ['python', '-m', 'pytest', '.']
=========================================================== test session starts ===========================================================
platform darwin -- Python 3.12.9, pytest-8.3.5, pluggy-1.5.0
rootdir: /Users/zhiqiu/offline_code/opensource/einops
configfile: pyproject.toml
collected 64 items                                                                                                                        

test_einsum.py ....                                                                                                                 [  6%]
test_examples.py .s                                                                                                                 [  9%]
test_layers.py ....ssss..                                                                                                           [ 25%]
test_ops.py ......................s.                                                                                                [ 62%]
test_other.py ............ss.s                                                                                                      [ 87%]
test_packing.py ....                                                                                                                [ 93%]
test_parsing.py ....                                                                                                                [100%]

============================================================ warnings summary =============================================================
einops/tests/test_layers.py::test_einmix_decomposition
  /Users/zhiqiu/offline_code/opensource/einops/einops/layers/_einmix.py:63: UserWarning: EinMix: weight has no dimensions (means multiplication by a number)
    self.initialize_einmix(

einops/tests/test_other.py::testmod
  /Users/zhiqiu/offline_code/opensource/einops/.venv/lib/python3.12/site-packages/_pytest/python.py:163: PytestReturnNotNoneWarning: Expected None, but einops/tests/test_other.py::testmod returned TestResults(failed=0, attempted=0), which will be an error in a future version of pytest.  Did you mean to use `assert` instead of `return`?
    warnings.warn(

-- Docs: https://docs.pytest.org/en/stable/how-to/capture-warnings.html
================================================ 55 passed, 9 skipped, 2 warnings in 0.18s ================================================

[ifsheldon]

Probably we need to add new CI configurations for testing MLX as well.

[ifsheldon]

@arogozhnikov Hi! Can you take a look at this PR? Thanks!

[Datamance]

Very excited for this! I prefer the MLX api and throwing einsum in the mix would be fabulous.

[ryang-max]

Any progress on this? Really looking forward to this feature 🚀

[Datamance]

When will this be merged?

[Datamance]

Ping 😄 @arogozhnikov can this be merged?

[arogozhnikov]

uh. Okok, let's make an exception for MLX.