This was inconsistently applied, with some types having <T>, some having
<T: Scalar>, and some having <T: RealField>.
This unifies all types to match the convention of Matrix:
Just declare <T> at type def time, and apply bounds on impls only.

A significant advantage of this approach is const fn construction. Const
fn generics currently still can't have trait bounds, so any generic
const fn needs to only move opaque types around. Construction methods
such as new_unchecked or from_parts can be made const by removing their
generic bounds after this PR.

Actual constification is left to a follow-up PR.

Note that na::Transform is _not_ loosened here, as it has more complicated
definition requirements.