Define values for ArraySpace

[jishnub]

Fixes #647.

values(f) errored for an array-valued Fun, so iszero broke for these
after #635 started consulting values:

julia> Af = Fun(t -> [0 1; -10cos(t)-1 -24-19sin(t)], Fourier(0..2pi));

julia> iszero(Af)
ERROR: Override for ApproxFunBase.ArraySpace(...)

There were three separate causes:

• itransform was only defined for a VectorSpace, so a MatrixSpace fell
  through to the generic Space method and reached checkcanonicalspace,
  which errors because canonicalspace is applied component-wise and an
  ArraySpace of canonical spaces is therefore its own canonical space. The
  method is now defined for an ArraySpace, and reshapes the values to the
  shape of the space.

• The component coefficients are views if the interlacer is trivial, in which
  case pad! can't resize them. This broke values for a VectorSpace as
  well, e.g. Fun(x->[exp(x),sin(x)], -1..1). pad is used instead.

• _values asserted a Vector{float(T)} return type for every space over a
  numeric domain. UnivariateSpace only constrains the domain type, so an
  ArraySpace over one matched, and the assertion rejected the array-valued
  result. The assertion is now restricted to spaces with a scalar range type.

Additionally, points(f::Fun{<:ArraySpace}) passed ncoefficients(f), the
total number of interlaced coefficients, where points(::ArraySpace, n)
expects a per-component count. This was silently wrong for Fourier, and an
error for a finite-dimensional component space such as PointSpace. It now
matches the padding used in itransform, so values(f) == f.(points(f)).

Tests cover PointSpace and HeavisideSpace components at vector and matrix
shapes; the latter is over a numeric domain and has a trivial interlacer, so
it exercises the second and third points above.

Downstream ApproxFun tests pass against this branch.

🤖 Generated with Claude Code