It allows removing an ugly input dependency in aout as well as finer control
at decoder side.

It allows replacing --codec dummy --dummy-save-es by --codec dump.

For lack of a better solution, I assume that Windows has a finite
paths length limit to prevent infinite recursion. Otherwise we need
dirfd() and unique inodes.

Forward port: [1a5a6b89692f2e6761b6ecd9ae9ee6e50b6e4138]

(cherry picked from commit b0438df5b9a5a7976be069ec173bb2077f97fd58)

(cherry picked from commit 31a711e42ea80553aca57626903cdd822c8c1bea)

(Not that I expect this to work at run-time)

forwardport [9439810b]

The core currently cannot handle a target name

We now account for this and use an "OpenPanel" instead of a "SavePanel".

Disconnects http-connection before trying with authentication info.
This enables to use http-authentication sources without leaking fds

Kinda work-around

Though the values are the same, this is cleaner, and more 'future' proof

This fixes a busy loop.