Implement `mapslices` without scalar iteration

When I use `mapslices(f,a,dims)` to manipulate CuArray, a warning appears. It reminds me that using scalar operations on the GPU is inefficient.

```julia
a=CUDA.rand(3,4,5)
b=CUDA.rand(2,3)
mapslices(a,dims=[1,2])do t
b*t
end
```
I had to use additional code to perform the operation.
```julia
c=map(eachslice(a,dims=3)) do t
b*t
end
cat(c...,dims=3)
```
In neural networks or machine learning, mini-batch is often used. When a sample is not a vector or matrix, the input of the model will have multiple dimensions each time, such as size(x)==(100,3,4,batch_size). In this case, `mapslices()` seems very convenient.

However, when the model and input are both CuArray, the GPU will be very inefficient due to too many scalar operations. Can the internal operations of `mapslices()` be optimized to make it more efficient?
**Describe the solution you'd like**

Is there a more elegant way to implement mapslices(f,a,dims) that enables it to use vectorized operations instead of scalar operations.