Type conversion in inits is not working as intended

[MartinuzziFrancesco]

Describe the bug

Type conversion in the initializers doesn't work as expected. It is fine for default arguments, but fails when user gives ie a Float64 as scaling while the matrix is supposed to be Float32.

Expected behavior

Matrix type should not be changed by the weight/scaling/other keyword optional changes made by the user.

Minimal Reproducible Example

julia> sr = scaled_rand(5, 5) # all good here, default kwargs 
5×5 Matrix{Float32}:
 -0.0669356   0.0944272  -0.0191563   0.0934286   0.0190222
  0.0159724  -0.0799005   0.0279123  -0.0188943   0.0148647
  0.026355   -0.0970182   0.0567368  -0.0737949  -0.0444782
 -0.0177412  -0.0292692   0.0679244   0.0714962   0.03553
 -0.0184405   0.004071   -0.0891089   0.0892906  -0.0399615

julia> sr = scaled_rand(5, 5; scaling = 0.5) #user gives a Float64 as kwarg, but matrix should still be Float32 (default value)
5×5 Matrix{Float64}:
 -0.334678    0.472136  -0.0957814   0.467143    0.0951111
  0.0798621  -0.399503   0.139562   -0.0944713   0.0743235
  0.131775   -0.485091   0.283684   -0.368974   -0.222391
 -0.0887059  -0.146346   0.339622    0.357481    0.17765
 -0.0922025   0.020355  -0.445545    0.446453   -0.199807

Potential solution

Just change type conversion instead of being here to be directly on the computations (here)