guard mlx_array_dim against 0-dim and out-of-bounds access

[TheTom]

Proposed changes

mlx_array_dim crashes with SmallVector out of range when called on
scalar (0-dim) arrays or with dim >= ndim. The exception is caught by
mlx_error which calls Swift's fatalError, killing the process.

This happens during Swift metadata initialization when Module subclass
hierarchies trigger lazy evaluation of 0-dimensional arrays. The crash
is non-deterministic and depends on binary layout, making it difficult
to reproduce but fatal when it hits.

Fix: guard against invalid dimensions before accessing shape().
Returns 1 for scalars, 0 for out-of-bounds. No change for valid inputs.

Repro

Any Swift binary linking MLXLMCommon (from mlx-swift-lm) with certain
class hierarchies can trigger this at load time:

MLX/ErrorHandler.swift:343: Fatal error: SmallVector out of range.
at mlx-c/mlx/c/array.cpp:335

Checklist

• I have read the CONTRIBUTING document
• I have run pre-commit run --all-files to format my code
• I have added tests that prove my fix is effective or that my feature works
• I have updated the necessary documentation (if needed)