ndarray has a history of encoding arrays’ dimensionality in the type system; this turns out to be useful both for debugging and for writing complex array libraries. However, some arrays don’t (or can’t) have their dimensionality known at compile time. One good example of this is the output of ArrayBase::squeeze: even if the array’s dimensionality is known before the operation, the dimensionality of the output depends on how many axes have length one.

The Dimensionality trait is intended to unify both the known- and unknown-dimensionality cases. Compile-time dimensionalities are represented using NDim, while cases where the dimensionality cannot be known at compile time are represented using DDyn.

A key design choice is that there is no way to recover the number of dimensions as a runtime value from a type implementing Dimensionality. A dynamic dimensionality here does not mean “known at runtime”, but rather “cannot be known at compile time”. The actual number of axes is always taken from the array’s shape, which avoids having two separate sources of truth.

…ions (#1575)

This is the next PR to address #1506. In the previous PR (#1568) we
established a trait for capturing the "dimensionality" or "rank" of an
array at the type level. This PR is focused on providing a bridge from
type-level dimensionality to runtime dimensionality, which we do through
the new `Ranked` trait.

The `Ranked` trait is pretty simple: require an associated type (`NDim`)
that carries compile-time dimensionality, and a function `fn ndim(&self)
-> usize` that tells the user runtime dimensionality. The
`src/layout/ranked.rs` file contains the definition, implementations,
and blanket implementations.

This commit also does some moves and renames of the existing
`Dimensionality` trait, renaming it to `Rank`, as well as many of
the types in that file.

CI on `refactor` is currently red because of newer toolchains, not code
changes. Since CI runs with `-D warnings`, the following are now errors:

- clippy 1.99 removed `clippy::from_iter_instead_of_collect`, so our
`allow` for it in `lib.rs` now errors.
- clippy 1.99 `question_mark` fires in `Indices` and the
`Iter`/`IndexedIter` iterators.
- clippy 1.99 `manual_checked_ops` fires in `into_scalar`.
- Nightly (and stable in the weekly latest-deps job) now deprecates the
legacy `std::f64` / `std::isize` modules and constants. These are used
in `tests/numeric.rs` and in tests in `dimension/mod.rs` and `tri.rs`.

No behavior change.

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>

#1575 moved `ndim` off of `Dimension` and onto the new `Ranked`
supertrait. Generic code with a `D: Dimension` bound still works, but
calling `.ndim()` on a concrete dimension type now needs `Ranked` in
scope, so this stopped compiling:

```rust
use ndarray::prelude::*;
Ix2(3, 4).ndim(); // E0599: trait `Ranked` is not in scope
```

This exports `Ranked` from the prelude next to `Dimension`. Code that
imports `ndarray::Dimension` directly, rather than going through the
prelude, still has to import `ndarray::layout::ranked::Ranked` too.

CI is red until #1622 lands. That PR fixes lints from newer toolchains
that also fail on `refactor` itself.

Co-authored-by: Devin AI <158243242+devin-ai-integration[bot]@users.noreply.github.com>