Reduction not found

Hi, here is a MWE of an unexpected error:

```
using LoopVectorization

function test(a, b)
#s = 0.0
@turbo for i in axes(a,1)
ai = 0.0
for j in axes(b,2)
bij = b[i, j]
ai += bij^2
end
ai = exp(ai)
a[i] = ai
end
return nothing
end

a = rand(256)
b = rand(256, 256)
test(a, b)
```

LoadError: LoadError: "Reduction not found."
Stacktrace:
[1] reduction_zero
@ $HOME\.julia\packages\LoopVectorization\7gWfp\src\modeling\costs.jl:649 [inlined]
[2] add_reduction_update_parent!(vparents::Vector{LoopVectorization.Operation}, deps::Vector{Symbol}, reduceddeps::Vector{Symbol}, ls::LoopVectorization.LoopSet, parent::LoopVectorization.Operation, instr::LoopVectorization.Instruction, reduction_ind::Int64, elementbytes::Int64)
@ LoopVectorization $HOME\.julia\packages\LoopVectorization\7gWfp\src\parse\add_compute.jl:255



If I use `a[i] = exp(ai)` instead of the last two lines, it works. It also works if the last two lines are `ai = exp(ai/2); a[i] = ai`. I feel the issue is related to your explanation in #498, but honestly I don't fully understand that. My general feeling there is that the error it gives is unnecessary and can be ignored if it can run... Is that correct?

Run the supplied Julia MWE with LoopVectorization and start by reading reduction_zero in src/modeling/costs.jl around line 649, then add_reduction_update_parent! in src/parse/add_compute.jl around line 255. Compare the failing assignment with the working variants and determine whether the reduction error is expected; done means the behavior is explained and the MWE no longer produces an unjustified error.

Written by the indexing model from the issue text.