`Bound<PyArray>>`, `PyReadonlyArray` and `PyReadwriteArray` all create useless `extract` messages on dtype/shape mismatch

[flying-sheep]

We should figure out how to use the very nice extract implementation that exists for PyArray instead of this; the cast here ignores the underlying extract error and just gives a generic (and in some cases nonsensical) error:

rust-numpy/src/borrow/mod.rs

Line 246 in 4ef0102

let array = obj.cast::<PyArray<T, D>>()?;

You removed the underlying detailed error from the public API in #315. We should use its information here.

Repro:

use numpy::PyReadonlyArray2;
use pyo3::prelude::*;
use pyo3::types::PyAnyMethods;

fn main() -> PyResult<()> {
    Python::attach(|py| {
        // A numpy.ndarray with the wrong dtype (float64 instead of float32)
        py.run(
            c"import numpy as np; arr = np.zeros((2, 2), dtype=np.float64)",
            None, None,
        )?;
        let arr = py.eval(c"arr", None, None)?;

        // Extracting as PyReadonlyArray2<f32> fails because the dtype doesn’t match
        let err = arr.extract::<PyReadonlyArray2<f32>>().unwrap_err();

        println!("arr’s Python type: {}", arr.get_type().name()?);
        println!("Error message:     {err}");
        
        Ok(())
    })
}

Outputs:

arr’s Python type: ndarray
Error message:     'ndarray' object is not an instance of 'ndarray'

Cargo.toml

[package]
name = "numpy-cast-error-mre"
version = "0.1.0"
edition = "2024"

[dependencies]
numpy = "0.29.0"
pyo3 = { version = "0.29.0", features = ["auto-initialize"] }

[Icxolu]

Yeah, we had this come up multiple times already in the past. The error message is a quite unfortunate. It's a combination of the generics used for type checking within numpy and having to go through pyo3s type checking mechanism which is not really build for this. The produced error and error message are from within pyo3 and numpy can't adapt them. We actually have the relevant information here

rust-numpy/src/array.rs

Lines 144 to 149 in 4ef0102

	fn extract<'a, 'py, E>(
	ob: &'a Bound<'py, PyAny>,
	check: unsafe fn(Python<'py>, *mut ffi::PyObject) -> c_int,
	) -> Result<&'a Bound<'py, Self>, E>
	where
	E: From<CastError<'a, 'py>> + From<DimensionalityError> + From<TypeError<'py>>,

but there is not real way for us to surface it to the user, which is why we currently just discard it.

[flying-sheep]

But we don’t have to, we can just call our “extract” method no? I made a PR to illustrate