Refactor trajectory generation to separate process

This pulls all the trajectory planning into a TrajectoryGenerator class,
which produces a Trajectory spline that the drivetrain code can consume
and use to track the spline.

Broadly speaking, this change:
- Separates the Trajectory class into a generation class and a
  FinishedTrajectory class, where the generator produces a flatbuffer
  and the FinishedTrajectory reads all the required information from
  the flatbuffer.
- Add an option for serialization/deserialization of a DistanceSpline.
- Removes some dead code from Trajectory class (mostly having to do with
  the old feedback algorithm).
- Uses floats in more places, to keep the size of the Trajectory
  flatbuffer under control
- Update the tests & autonomous code to use the new spline code.

Further work that may make sense:
- Experiment with alternatives to current structure of the Trajectory
  flatbuffer to see whether (a) the size is a problem; and (b) if so,
  what we should do about it.
- Add shims to allow replaying logfiles with old-style spline goals.

Change-Id: Ic80ce4e384ec4d1bd22940580e3652ecd305b352
diff --git a/y2020/BUILD b/y2020/BUILD
index 024b6ac..90d976b 100644
--- a/y2020/BUILD
+++ b/y2020/BUILD
@@ -12,7 +12,7 @@
         ":config",
     ],
     dirs = [
-      "//y2020/actors:splines"
+        "//y2020/actors:splines",
     ],
     start_binaries = [
         "//aos/events/logging:logger_main",
@@ -22,6 +22,7 @@
         "//aos/network:message_bridge_server",
         "//y2020/actors:binaries",
         "//y2020/control_loops/drivetrain:drivetrain",
+        "//y2020/control_loops/drivetrain:trajectory_generator",
         "//y2020/control_loops/superstructure:superstructure",
     ],
     target_compatible_with = ["@platforms//os:linux"],