Proof-of-concept implementation of Static Structural

[saeta]

Based on explorations of structural generic programming in
tensorflow/swift-models#650 we would like to support
smooth interactions between KeyPaths and structural generic programming.
Based on discussions with @shabalind, I have put together a quick
proof-of-concept demonstrating how to unify a "static structural"
implementation with the "instance"-level structural generic programming we
have been exploring so far.

This is definitely an incomplete implementation, but this would allow us to
easily implement KeyPathIterable and its ilk on top of structural generic
programming (something not possible previously), in addition to unifying with
the existing KeyPath world.

There are a couple further extensions:

1. Solution for HNil To use a proper cons-list construction, this will
   require a bottom type (Never) that conforms to every protocol, and
   multiple conformances. Alternatively, an encoding that appears to work
   is a cons-list construction that doesn't use a special end type.
2. Typealias inference: Swift does not currently infer typealiases for
   these kinds of derived conformances.
3. Enum support This current implementation doesn't include enum support
   although this is straight forward to add.
4. Simplier structural representations: When the single static structural
   derived conformance, we can easily implement multiple simpler "structural"
   encodings (e.g. a simple Cons-list of values in addition to the current
   structural representation).
5. Extensions to KeyPaths could enable StaticStructural to simply be the key
   paths themselves. (i.e. if you could retrieve the field name corresponding
   to the key path.)

[saeta]

CC @shabalind @dabrahams @dan-zheng

[shabalind]

Closing this off for now similarly to #10. We might revisit it later.