Implement Sylvester NFs (van den Berg et al. 2019) with a tanh residual transform, and Planar as a special case

[bhvieira]

This is the first attempt to add Sylvester Normalizing Flows (van den Berg et al. 2019), and also Planar flows (Rezende et al. 2015), both listed in #28 as requested flow architectures.

The inverse and thus sampling are solved through a bisection method.

I'm open to comments and suggestions! Let me know if anything needs improvement. The part I'm least sure about is the inverse:

• The planar transform’s denominator clamp breaks its invertibility guarantee
• Small norm householder vectors produce a nonorthogonal matrix
• Float32 saturation in tanh lets diagonal products reach -1 producing an infinite negative log determinant
  I haven't found these issues in practice yet, but I think the code is definitely vulnerable to these