Add metadata to exported functions

[eyupcanakman]

Closes #2410. Exported .mlxfn files can now store a metadata dictionary, so a model's config can live with the function instead of in a separate JSON file.

export_function and exporter take an optional metadata argument, a dict of string keys to string, int, or float values. import_function(file, return_metadata=True) returns the function along with that dict, mirroring load(..., return_metadata=True) for safetensors and gguf. The value shape follows the one suggested in the issue.

Integer and float values are stored as 64-bit so a Python int or float round-trips without narrowing. Metadata is only supported when exporting to a file. Passing it with a callback raises.

The metadata lives in the file header, so a .mlxfn written before this change has no metadata section and won't load in a metadata-aware build. That matches the existing note that exported functions aren't compatible across MLX versions.