vmap gives incorrect results for iszero

If I run vmap with certain functions i get wrong results in some cases. I noticed this first for an integer array with iszero:
```
a = rand(Int8,200,20)
@test map(iszero,a) == vmap(iszero,a) #fails

```
Repeated evaluations of `vmap(iszero,a)` give different answers each time.
The same happens for some other functions I tested such as:

```
a = rand(200,20) #also for floats
vmap( ==(2.0), a)
vmap( >(0.5), a)
```
As a pattern these all return boolean arrays, functions that return floats seem to be unaffected.

vmapreduce is unaffected:
```
a = rand( Int8,200,20)
@test vmapreduce(iszero,+,a) == mapreduce(iszero,+,a) #passes
@test sum(map(iszero,a)) == sum(vmap(iszero,a)) #fails
```
The inplace version seems affected as well:
```
a = rand( Int8,200,20)
out = rand( Bool,200,20)
vmap!(iszero,out,a)
@test out == map(iszero,a) #fails
```
Interestingly this seems to consistently return the same (wrong) result instead of the always changing values of the out of place version. Here again this seems to only affect functions that return bools.

This is just from some playing around not something I actually need but unless I'm violating some assumptions LoopVectorization makes this seems like a bug.

Reproduisez l’écart avec les exemples fournis `vmap(iszero, a)` et `vmap!`, en comparant chaque résultat avec `map` ; vérifiez également que `vmapreduce` reste correct. Commencez par les points d’entrée `vmap` et `vmap!` et suivez les fonctions qui renvoient des valeurs booléennes, puis ajoutez une couverture de régression montrant des résultats stables égaux à ceux de `map` pour les cas signalés.

Rédigé par le modèle d'indexation à partir du texte de l'issue.