Improve stride copying in generalize a bit

While still not perfect - that should be solved in upstream ndarray, the
new version is an improvement since it checks the compatibility before
attempting to construct the array.

`.into_raw_vec()` is very hard to use correctly (this is inherent, it's
an access to the raw memory model), so the new version will actually
fail to convert in some cases that were passing as silent errors before
(related to internally sliced arrays, rather uncommon).

ndarray-linalg/src/convert.rs

@@ -102,18 +102,21 @@ where
 {
  // FIXME
  // https:/bluss/rust-ndarray/issues/325
- let strides: Vec<isize> = a.strides().to_vec();
- let new = if a.is_standard_layout() {
- ArrayBase::from_shape_vec(a.dim(), a.into_raw_vec()).unwrap()
- } else {
- ArrayBase::from_shape_vec(a.dim().f(), a.into_raw_vec()).unwrap()
- };
+ //
+ // copy strides
+ let mut strides = D::zeros(a.ndim());
+ for (index, &s) in a.strides().iter().enumerate() {
+ strides[index] = s as usize;
+ }
+ let a_dim = a.raw_dim();
+ let a_len = a.len();
+ let data = a.into_raw_vec();
  assert_eq!(
- new.strides(),
- strides.as_slice(),
- "Custom stride is not supported"
+ a_len,
+ data.len(),
+ "generalize: non-contig arrays are not supported"
  );
- new
+ ArrayBase::from_shape_vec(a_dim.strides(strides), data).unwrap()
 }
 
 /// Fills in the remainder of a Hermitian matrix that's represented by only one