This builds upon the refinement introduced by:
#1437

That PR allowed us to use a good packet ignoring skips if the
previous callback was serviced using a prediction (glitch).

However, we were still effectively throwing out the next good
packet whenever there were skips and the previous callback was
not serviced using a prediction.

This PR takes a slightly different approach by allowing us to
stash and use the next good packet on the following callback.
This also helps slow down jumping sequence numbers when there
are multiple skipped packets in a row (which causes abrupt
drops in latency).

This also addresses a special case where we are only skipping
a single good packet due to tolerance. Rather than returning
a prediction, we allow it to use the packet.

For this to work, we need to ease the tolerance restrictions
slightly by allowing it to sometimes expand by the duration
of an additional packet. Effectively, this also means that
tolerance restrictions alone will only cause glitches when
the skew exceeds at least one packet in duration.

My initial testing indicates that these changes significantly
reduce glitch counts, by as much as 2x, resulting in a
slightly improved signal (at least to my ears).

sudo tc qdisc add dev eth0 root netem slot distribution pareto 0.1ms 3.0ms loss 10%

This also eliminates the use of "goto" statements in the
regulator code and fixes a bug with auto headroom adjustments