This allows more flexibility in processing the animations before
serialization. When the track time is present, we write it instead of a
fixed-rate frequency track.

For now no attempt to deduplicate time tracks is taken; this could be
added in the future if need be, but the assumption is that the more
optimal way to serialize tracks with shared time ranges is to use
resampling anyway.

Tracks with cubic splines have three output components for each input
component; also, they require writing interpolation type properly.

When serializing cubic spline tracks for rotations, the tangents are not
actually quaternions; as such, they can't be encoded using quaternion
filter, and in -cc mode we need to fall back to raw snorm encoding. Even
snorm encoding technically might not be sufficient as tangents could be
outside of -1..1 range, but for now we can probably clamp out of range
tangents.

Using -af 0 now disables resampling and keeps the original time values.

When resampling is disabled, we can still mark tracks as constant and
remove them if they are not necessary, which cuts down on the animation
size. In this mode it is more difficult to support ranges, so we assume track
range will cover the relevant timespan if any tracks with time survive.

Doing this analysis for cubic splines is slightly more difficult as if
tangents are non-zero, an otherwise "constant" track will not result in
constant output, so we disable analysis in that case for simplicity.

When tracks have custom time inputs (and resampling is disabled), we can
reuse inputs when they are equal between two tracks. This sharing is
only performed between consecutive tracks as it allows us to still
process tracks linearly.