Possibility of doing all computations using `Float32` for performance

This is some notes from my discussions with @santiagobadia and @amartinhuertas, regarding the option to use `Float32` everywhere. Currently there are places where `Float64` is hard-coded, especially the quadrature.

- For example, we can see in `/src/ReferenceFEs/TensorProductQuadratures.jl`, that `T=Float64` is hard-coded into `_tensor_product_legendre(degrees)` function. (`gauss` quadrature from `QuadGK.jl`, which is used here, can handle `Float32`)
- Both `CartesianDiscreteModel{D,T,F}` and `UnstructuredDiscreteModel{Dc,Dp,Tp,B}` carry Floating-point type parameter `T` and `Tp` respectively, which can be propagated into quadrature constructors. But we note that their supertype `DiscreteModel` doesn't have this type parameter
- Even if the model itself is in `Float64`, we can have the option of quadrature to be `Float32`, in cases where we are not using FE basis but only using `Gridap` as means to integrate some external functions.