Bug in BezierSpline

[bkp7]

There is a bug in the BezierSpline.predict code (no bug in either the Linear or Cubic versions). It fails if the target_x is of type s32. It works fine if the target_x is f32.

This variation on the documented example fails:

iex> x = Nx.iota({4})
iex> y = Nx.tensor([2.0, 0.0, 1.0, 0.5])
iex> model = Scholar.Interpolation.BezierSpline.fit(x, y)
iex> Scholar.Interpolation.BezierSpline.predict(model, Nx.tensor([3]))

[josevalim]

I believe this is fixed as of #311. Can you please try main?