Preserve element type in stateless layers

[staticfloat]

This should fix issues where Float32 weights and activations get suddenly switched to Float64 on the backwards pass. MWE to show current problem:

using Flux

x = param(randn(Float32, 10, 10, 4, 4)
mp = MeanPool((10,10))
Flux.back!(Flux.mse(mp(x), mp(x)))

This currently errors out with:

ERROR: MethodError: no method matching ∇meanpool(::Array{Float64,4}, ::Array{Float32,4}, ::Array{Float32,4}, ::Tuple{Int64,Int64}; pad=(0, 0), stride=(10, 10))

This is because the backward pass of mse() is spuriously generating Float64's. I don't really understand why it is doing that, but this PR seems to fix it.

[staticfloat]

It's possible this will fix #501

[DhairyaLGandhi]

For further reference, #514, #538, #533 might be relevant.

[KristofferC]

This is because the backward pass of mse() is spuriously generating Float64's. I don't really understand why it is doing that, but this PR seems to fix it.

It does so because it divides by an Int and inv(::Integer) gives a Float64 (in the backwards pass). I think #538 is a cleaner solution.

[staticfloat]

Agreed; #538 is the cleaner solution here.