Fix find() when called with function and non-StridedArray

[nalimilan]

The previous code expected testfun to be vectorized, which does not match
the behavior of other methods, nor the documentation. Especially visible
when passing a sparse matrix. Always return an Array rather than an object
similar to the input, because it makes no sense to return a sparse vector.
Remove the method matching any object so that an error is raised rather
than returning incorrect results on non-AbstractArray (e.g. Set).

This fixes the bug reported by i.pallikarakis-11 on the mailing list, namely:

A=speye(Bool,10)
find(x->x==true,A)
->0-element Array{Int64,1}

find(x->x==true,full(A))
->10-element Array{Int64,1}:
   1
  12
  23
  34
  45
  56
  67
  78
  89
 100

I'm not completely sure about the AbstractArray requirement, but it doesn't make sense to me to return indexes for non-arrays. The previous code would happily work on Sets, but would return nonsensical results.

(BTW, I'm aware that calling find(testf, x) with x a sparse matrix is not very efficient, but at least it must not give wrong results.)

[nalimilan]

I realize this change has the drawback that scalars are only supported if they are Numbers:

julia> find(s -> ismatch(r"a", s), ["az"])
1-element Array{Int64,1}:
 1

julia> find(s -> ismatch(r"a", s), "az")
ERROR: `find` has no method matching find(::Function, ::ASCIIString)

But if find works on any object, not only numbers, then it gives misleading results for non-scalars like Set.

[ivarne]

See also my answer to the other post about the problem.

This looks like a bug to me, and it was caused by c330b1d and it links to #987.

The bug is silent because there already was a generic function that could find things in non iterable collections introduced in b8f47e8 (that seems like a stupid use of find and it should probably be restricted to Number).

[JeffBezanson]

Should be !testf(x) instead of testf(x)==0, but false == 0 so this works.

[nalimilan]

Yeah, not pretty. :-/ I could have fixed that. Worth another commit?

[ivarne]

Not sure, but a commit is cheap: 6b3276f

[nalimilan]

Yes, but for me it implies bothering somebody with a pull request.

While we're tweaking style, how about writing it testf(x) ? [1] : Array(Int,0) instead? :-p

[ivarne]

Or

testf(x) ? Int[1] : Int[]

Which will also make the function type stable. (Sorry! Array constructors confuse me)

I don't think Pull requests is bothering for small changes like this, that are obviously correct. A unrelated discussion that goes on and on, is a different matter tough.

[rfourquet]

Or testf(x) ? [1] : Int[] ?!

[JeffBezanson]

cc @JuliaBackports

[ivarne]

Backported in f0214c3