restrict second argument of size and axes to Integer for most types

[JeffBezanson]

This argument had a variety of strange behaviors:

julia> size([1], Int8(1))
ERROR: TypeError: in arraysize, expected Int64, got Int8

julia> size((), Int8(1))
0

julia> size((), :x)
ERROR: ArgumentError: invalid tuple dimension x

julia> size((1,), 1.0)
1

julia> size(1, :x)
ERROR: MethodError: Cannot `convert` an object of type Symbol to an object of type Int64

After:

julia> size([1], Int8(1))
1

julia> size((), Int8(1))
0

julia> size((), :x)
ERROR: MethodError: no method matching size(::Tuple{}, ::Symbol)

julia> size((1,), 1.0)
ERROR: MethodError: no method matching size(::Tuple{Int64}, ::Float64)

julia> size(1, :x)
ERROR: MethodError: no method matching size(::Int64, ::Symbol)

[staticfloat]

This seems like a mostly-unrelated change; is this just cleanup? Do the type arguments help codegen because of @noinline?

[JeffBezanson]

Yes, it's unrelated. Just good style to write the types since we know them exactly.

[JeffBezanson]

Ok, this shouldn't be done for the AbstractArray fallback since it will introduce new ambiguities. Changed to a typeassert instead.