Add initial MLX support with the creation wrappers

[aaishwarymishra]

This pull request introduces initial support for the MLX backend in the array_api_compat package. The main changes add MLX-specific modules, implement compatibility wrappers for core array functions, and update the build and development configuration to include MLX. Below are the most important changes:

For #452

MLX Backend Implementation:

• Added the new array_api_compat/mlx backend with __init__.py, _aliases.py, and _typing.py to support MLX arrays and provide compatibility wrappers for common array operations.
• Implemented MLX-compatible versions of arange, asarray, empty_like, eye, meshgrid, ones_like, and zeros_like in _aliases.py, handling MLX-specific quirks and edge cases.
• Defined MLX-specific type aliases (Array, DType, Device) in _typing.py for consistent type handling.
• Updated __init__.py to expose all relevant symbols and ensure proper module structure for the MLX backend.

Build and Development Environment:

• Updated meson.build and pyproject.toml to include MLX sources and conditionally require the mlx package on macOS (sys_platform == 'darwin'). [1] [2]

[Copilot]

Copilot was unable to review this pull request because the user who requested the review has reached their quota limit.

[aaishwarymishra]

After adding temporary compatibility wrappers and rerunning test_creation_functions.py, I found a few additional MLX limitations:

• arange dtype: MLX defaults integer inputs to float32, while the Array API expects int32. The wrapper uses int32 for integer-only inputs and float32 otherwise.
• Large integer arange: mx.arange rejects some values above 2**31 - 1. Empty and single-element ranges can be handled, but nontrivial int64 and uint64 ranges still fail.
• eye with int64/uint64: MLX GPU scatter does not support these dtypes, so the wrapper creates the matrix in int32 and casts it.
• asarray(copy=False): Returning an existing compatible MLX array avoids a copy error, but the aliasing test fails because MLX does not support 0-D assignment with x[...] = value.
• empty_like, ones_like, zeros_like: Wrappers use x.dtype when dtype=None.
• meshgrid: The wrapper converts MLX’s list output to the required tuple.
• linspace: Still fails because MLX does not support endpoint. No wrapper was added due to possible numerical differences.

With:

ARRAY_API_TESTS_SKIP_DTYPES=float64,complex64,int64,uint64

the suite reports 14 passed, 2 failed:
Note: Some times we got lucky with arange and other methods due to hypothesis gives different values which sometimes passes the tests

1. test_asarray_arrays: unsupported 0-D Ellipsis assignment.
2. test_linspace: missing endpoint support.

NOTE: MLX does not implement eq for comparing None with datatype, also we skipped complex completely due to problem with complex() but tests doesn't allow completely skipping complex, I think tests can be improved here

[ev-br]

Thanks!

Going forward, it would be helpful to list reproducible examples from the test suite failures, which the test suite prints on a failure.

complex() but tests doesn't allow completely skipping complex, I think tests can be improved here

Please open an issue in the test suite, with an MRE. If it's a test suite problem, it should be reproducible with numpy or array-api-strict, I suppose.

MLX does not implement eq for comparing None with datatype

This is one of things to propose adding to MLX; even if they decline, we'll have an answer.
Similar for other fixes, large and small.