Add tests for LoopVectorization

[mcabbott]

Closes #51 , closes #38 .

Lines with avx=false in test/parsing.jl produce errors (LoopVectorization v0.9.1, Tullio 0.2.10), although many more don't work with LV (either caught by try/catch during expansion, or else never called by dispatch).

Tullio.onlyone is part of the vectorised gradient for reductions over max, not sure it's sufficiently tested. now tested, but gradients still seem problematic. But I don't think this is a regression.

7896059 and 90f32e3 (on master before this branch) are what's needed for LoopVectorization v0.9. first attempts.

[chriselrod]

Glad there was at least something that worked with LV.

    @test_broken [3,7] == let s = [1 2;3 4], zero=one # left because of #50
-       @tullio x[i] := s[i,j]
+       @tullio x[i] := s[i,j]  avx=false # Unexpected Pass with LV
    end

Feel free to file issues for all of them that didn't.
I'd be inclined to at least fix all of them related to VectorizationBase (seems to be a few, like type error with AbstractSIMD or method errors, and probably the stack overflows as well?), aiming for the others after the rewrite.

@tullio A[i] := (1:10)[i]^2  verbose=1 avx=false # @error: rejected by LoopVectorization's check_arg

That happened unintentionally when reimplementing check_arg.

[mcabbott]

Clearly your macro has better hygiene than mine :)

But sure, I will gather up a list of these & make an issue, once things are a little tidier.

[chriselrod]

But sure, I will gather up a list of these & make an issue, once things are a little tidier.

Okay. I could try running tests myself too, I guess, but for now I did add a trunc(::Type{<:Integer}, ::AbstractSIMD) method for the trunc test.