Clean up parallel adaptation code.

 - The AdaptiveTrackSelection doesn't need to use the experimental
terminolgy because the code is always triggered if there are multiple
adaptive selections.

- It's also confusing to pass the state on the outside after the object
creation, so moving everything into a simple control flow again where
the adaptation checkpoints are passed in via the constructor.

- Instead of triple arrays, we can use more readable named structures.

- The calculation of the checkpoints can be cleaned up to be more
readable by moving things to helper methods.

- The reserved bandwidth from all fixed track selections is really just
a special case of multiple parallel adaptataions. So this logic doesn't
need to be separate.

- The whole logic also didn't have test coverage so far. Added tests
for the actual adaptation using these checkpoints and the builder
calculating the checkpoints.

Overall this should be a no-op change.

PiperOrigin-RevId: 350162834