Ndonnx

------------------

**Bug fixes**

- Fixes parsing numpy arrays of type ``object`` (consisting of strings) as ``utf8``. Previously this worked correctly only for 1d arrays.

**Breaking change**

- :meth:`ndonnx.Array.shape` now strictly returns a ``tuple[int | None, ...]``, with unknown dimensions denoted by ``None``. This relies on ONNX shape inference for lazy arrays.

------------------

**New features**

- Expose the :func:`ndonnx.isdtype` function.
- Custom data types can now override array functions:
- :func:`ndonnx.zeros`
- :func:`ndonnx.zeros_like`
- :func:`ndonnx.ones`
- :func:`ndonnx.ones_like`
- :func:`ndonnx.full`
- :func:`ndonnx.full_like`
- :func:`ndonnx.arange`
- :func:`ndonnx.arange`
- :func:`ndonnx.eye`
- :func:`ndonnx.tril`
- :func:`ndonnx.triu`
- :func:`ndonnx.linspace`
- :func:`ndonnx.where`
- The :class:`ndonnx._experimental.UniformShapeOperations` now provides implementations of shape operations that are generic across all data types where each constituent field has the same shape (that of the overall array).

**Other changes**

- Fixed various deprecation warnings.
- Invoking a function using arrays with data types that lack a corresponding implementation now raise a :class:`UnsupportedOperationError`.

**Bug fixes**

- Numerical operations like :func:`sin` now raise :class:`UnsupportedOperationError` when invoked using invalid data types like ``ndx.utf8`` rather than implicitly casting.
- Fixes bug causing a promotion error when implementing numerical operations like :func:`add` that involve type promotion.
- Fixes scalar promotion logic to more accurately reflect the Array API standard. Promotion requires at least one array to be present and scalars adopt the dtype of the arrays being promoted with it. `ndx.utf8` and `ndx.nutf8` cannot be promoted with any other dtypes.
- Fixes failure when broadcasting nullable data type arrays together in :func:`broadcast_arrays`.

------------------

**Bug fixes**

- Division now complies more strictly with the Array API standard by returning a floating-point result regardless of input data types.

------------------

**Other changes**

- ``ndonnx.promote_nullable`` is now publicly exported.

------------------

**Other changes**

- ndonnx now exports type annotations.

**Bug fixes**

- ``__array_namespace__`` now accepts the optional ``api_version`` argument to specify the version of the Array API to use.

------------------

**Breaking changes**

- The constant propagated value is no longer accessed from the ``eager_value`` property but instead the ``to_numpy()`` method.
- Non Array API functions have been moved to the ``ndonnx.additional`` namespace.