Read first before using snd_pcm_wait(), just like arecord does.
(cherry picked from commit 4b0e0dcd)

i_audio_mode should be used for the audio input mode (if ever supported).

Also do not enumerate every tuner two times when one is enough.

The table is not needed. Furthermore, each input is now enumerated
once rather than twice.

Maybe one day we will support autodetecting V4L2 device nodes
(as file://) but until then...

For the time being, mmap() is used if supported, otherwise read() if
supported, otherwise fail. This makes the user pointer method dead
code. It might make sense to revive it on capable device drivers to
support zero copy, but until then that method is no better than mmap().

Also an integer values in the MRL (v4l2://standard=XXX) is now
interpreted as a v4l2_std_id bit field, just like the command line
option (--v4l2-standard=XXX). This fixes an out-of-bound read with
large values.

Keeping the list of video standards in memory was almost pure waste:
The whole data set was only used to print one debug message.

The list of supported standards depends on the selected input.
So listing the standards during probe was wrong in any case.

Oops :D

This should fix compilation with kernel headers 2.6.27 and older
(obviously not tested).

libv4l2 directly wraps the system call if it does not know the file
descriptor. There is no need for VLC to reinvent the wheel.

"libv4l *always* reports emulated formats to ENUM_FMT,
 except when conversion is disabled."

When libv4l2 is used:
 * open the device with close-on-exec descriptor flag,
 * convert from UTF-8 to local path.