Fix tests for current Julia nightly

[martinholters]

The first commit replaces calls to CartesianIndices(1:3, 1:2) with CartesianIndices((Base.OneTo(3), Base.OneTo(2))). The former now gives an error when used for indexing, e.g.:

julia> zeros(3, 2)[CartesianIndices((1:3, 1:2))]
ERROR: MethodError: no method matching similar(::Array{Float64,2}, ::Type{Float64}, ::Tuple{UnitRange{Int64},UnitRange{Int64}})

Is that supposed to happen, @mbauman?

The second commit removes the testfindfirst on IOBuffer as Julia removed support for it in JuliaLang/julia#26600. Should we also add a deprecated method to findfirst, using the definition of the deprecated search, i.e. https://github.com/JuliaLang/julia/blob/278a2893b8b02d6063ee4b1717c201c3431cb042/base/deprecated.jl#L1565-L1571 ?

[martinholters]

Looks like I should have checked on current current Julia nightly... Will look into the remaining failures.

Meanwhile CC @nalimilan for the findfirst thing.

[nalimilan]

Thanks. I don't think we should add a deprecation to findfirst given that we only define the method on Julia 0.6. If we deprecated it there packages would start printing annoying warnings even for users who don't care about 0.7.

Regarding CartesianIndices, why not use CartesianIndices((1:3, 1:3)) instead of OneTo?

[martinholters]

Regarding CartesianIndices, why not use CartesianIndices((1:3, 1:3)) instead of OneTo?

JuliaLang/julia#26682 (comment)

[martinholters]

The remaining failure is not just a problem with the tests. So I propose we merge this as is, which will enable a future PR fixing round and friends to actually see its tests run and get a green CI.

[nalimilan]

I'd rather wait until we have a clearer view of what we want to do, or disable/remove the map(i -> l[i], c) test, as I'm not sure it's supposed to work at all. I've added it when backporting CartesianIndices, but it's not tested in Base obviously, so it's fine if Compat doesn't test it either (at least for now). Using Base.OneTo makes less sense to me since we mainly want to test ordinary ranges, which users are more likely to use.

[martinholters]

I'd rather wait until we have a clearer view of what we want to do, or disable/remove the map(i -> l[i], c) test

Yep, will do. But note that it's also the l[c] that fails.

[martinholters]

Ah, test failures on the out-of-data Julia nightlies are due to a mistake in #530. I'll add the fix to this PR to verify I get CI green...

[martinholters]

Good to go?