Only on fast seeking sources

Close #12724
Signed-off-by: Jean-Baptiste Kempf <jb@videolan.org>

Signed-off-by: Jean-Baptiste Kempf <jb@videolan.org>

fixes #13967
* Recenters everything that depends on program: Time, timerolls, hacks...
* Removes PCR probing, programless timeroll handling
with a simplier fix (won't work with 2x26 hours recording, but..)
* Better probing of program duration (not based on inter PCR). More accurate,
at best by 1 frame length.
* Changes seek handling and bisects only on request (removes the whole file probing)
* No longer creates ES at preparse time
* Simplify/Unify dual program selection list and don't trust default program at first
* Many numerous other fixes

Moves the PCR fix from a global to program focused one.

Replaces the program PCR with a selected one instead of
sending PCR updates from all streams (applies to --no-trust-pcr)

Queues blocks until PCR appears. (will help fixing offsets)

Changes PCR probing time from 100ms (spec max) to 0.5s
(seen some streams with 0.4s inter-PCR)

Truncated packets at the end

Now we no longer update times and statistics in the control loop
(where it was kinda pointless). If there is no wake-up from the ES
output, then the input thread only needs to wait for control requests
- which means it can sleep without time-out.

In practice, that corresponds to the input thread being paused and
not buffering (buffering while paused is possible due to seek).

Before the next patch, the deadline should always be zero or positive.
(Also note that deadlines, unlike time-outs are absolute timestamps.)

The values will typically not change while waiting for control requests
and not processing the demux and outputs. Even if the values do change
(e.g. due to a seek request), the demux needs to be called to recompute
the updated values: see demux_Control() calls in MainLoopStatistics().

Note that the intrinsic computational savings from this commit are
probably negligible. The interesting aspect is the removal of one
wake-up source within the input control loop. This commit does not
remove nor add any wake-up however.

This is always VLC_SUCCESS anyway.

This is pending fixing or removal decision.

Consistent with historical semantics, the destination format will
be incomplete but valid if the copy fails. It must be cleared with
es_format_Clean() even in case of error (as before).

Also add a few assertions against invalid formats.

This is always VLC_SUCCESS, or a forwarded value. Ultimately, the
value was (rightfully) ignored by the stream output core.

Example:
 #cycle{duration=20m,
        dst=std{mux=ts,access=file,dst=sport.ts},   duration=5m},
        dst=std{mux=ts,access=file,dst=weather.ts}, duration=5m}

Skips 20 minutes, then records 5 minutes to sport.ts, then 5 more
minutes to weather.ts and restarts.

"duration" specifies the duration of the previous phase
"offset"   specifies the offset at which the previous phase ends
           and the next phase begins
           (mutually exclusive with duration)
"dst"      specifies the stream output chain for the phase
           (if missing, the phase is skipped/discarded)

Durations and offsets are so far expressed as an integer, optionally
followed by a unit: w=week, d=day, h=hour, m=minute, s=second. Second
is the default.

Currently only the decoding time stamp can be used as a reference, but
adding local or UTC clocks should be relatively easy.

ES synchronization and reference frames management is left for
further study.