automatically convert input matrix to Float32

[tiemvanderdeure]

This follows up on JuliaAI/MLJModels.jl#565 and just adds a tiny step that automatically converts input to Float32 before passing it to Flux.

I can't see why anyone would ever not want this, as the eltype of any neural net generated through MLJFlux will always be Float32, and any other input type will be converted to Float32 anyways, but with a much bigger computational cost. So I didn't build in an option to disable this behaviour.

The reason to have this in MLJFlux in particular is that other machines such as MLJModels.OneHotEncoder output Float64 types.

I didn't look into #267 in detail and it might make this redundant in some cases, but maybe not in all?

[ablaom]

Thanks indeed for this PR. Can you please look at the CI fails? For example, we have:

collate: Error During Test at /home/runner/work/MLJFlux.jl/MLJFlux.jl/test/core.jl:25
[648](https://github.com/FluxML/MLJFlux.jl/actions/runs/10738332003/job/29781691377?pr=272#step:6:651)
  Test threw exception
[649](https://github.com/FluxML/MLJFlux.jl/actions/runs/10738332003/job/29781691377?pr=272#step:6:652)
  Expression: MLJFlux.collate(model, X, y) == ([(Xmatrix')[:, 1:3], (Xmatrix')[:, 4:6], (Xmatrix')[:, 7:9], (Xmatrix')[:, 10:10]], rowvec.([y[1:3], y[4:6], y[7:9], y[10:10]]))
[650](https://github.com/FluxML/MLJFlux.jl/actions/runs/10738332003/job/29781691377?pr=272#step:6:653)
  MethodError: no method matching collate(::MLJFlux.NeuralNetworkRegressor{MLJFlux.Linear, Optimisers.Adam, typeof(Flux.Losses.mse)}, ::Tables.MatrixTable{Matrix{Float64}}, ::Vector{Float32})
[651](https://github.com/FluxML/MLJFlux.jl/actions/runs/10738332003/job/29781691377?pr=272#step:6:654)

[ablaom]

@tiemvanderdeure If you can fix the fails here, I can sort out the conflict.

[tiemvanderdeure]

Works now!

[ablaom]

Okay @tiemvanderdeure, thanks.

Turns out conflict resolution is a bit more complicated than anticipated. So I'm waiting on resolution of the following before trying again:

• Entity embedding layers should output Float32 not Float64, as present? #273

[ablaom]

@tiemvanderdeure Your PR has now been merged and is part of 0.6, just released.

Thanks for you patience as we sorted out the conflicts.

[tiemvanderdeure]

Awesome, thanks for seeing this one through!