Compile featurizers as part of serving computation

[jonatanklosko]

We can define featurizers in two chunks, an initial chunk that takes arbitrarily sized input of any allowed type and converts it to a normalized-size tensor, then an optional Nx chunk that does further processing. This plus some shape information should allow us to compile the Nx chunk together with the model in serving implementations.

[josevalim]

We should also include post processing in the computation too.