Since our SoXR audio filter use the soxr_clear function that is broken before
0.1.2.

The SoXR variable-rate engine is slower than the fixed one, so use it only when
it's needed (when the input ratio is changing to catch up a delay).

Keep it simpler, it shouldn't happen since we alloc an output buffer that is
slightly bigger than expected (in order to have enough room for data that is
held by SoxR).

These function will drain/flush the chain of filters of the audio output.

pf_audio_flush will flush/reset the state of an audio filter.
pf_audio_drain will drain an audio filter, it'll returns a block_t containing
the drained output.

Deprecated by adaptative, which is also now jammed by filter.
Drops specific MP4 forged smoo/stra hacks.

uses xml reader for encoding proofness
now reuses and shares xml reader

The PRI macros are used in some of the translated strings, so libintl
needs to support them.

Note that need-formatstring-macros nonobviously implies need-ngettext.

And avoids creating stats from non http chunks (smooth)

delete connection manager last,
streams will want to close unfinished downloads

Some smooth streaming implicit trun data offset are
pointing to (seems) tfhd 1st byte, instead of moof,
resulting of first samples pointing inside moof.

Better than late in-demuxer fix.
reverts 8b269f94

This reverts commit 4aeccbd7.

The decoder thread no longer needs the lock to use the aout, only to
modify the aout pointer.

This removes the last direct (ab)use of the audio output object from
the input thread, and the penultimate one of the video output object.

This solves some race conditions whereby the output pause state was
enabled by the input thread, while the decoder thread had decoded data
blocks to queue for playing. Decoded blocks should never get queued
during pause.

There are no real functional changes here. At the first iteration,
wait_acknowledge is no longer signaled upfront. Because the thread made
no observable changes by that point, the signal had no effects anyway.