From the CURLOPT_RANGE documentation: "The application does not have
to keep the string around after setting this option."

Turns out that using CP_ACP is a lousy idea, because only very few
Unicode characters can be represented by it.  Instead, switch to UTF-8
(which every sane person on other operating system already uses).

Closes #102

Applies the improvements from the previous commit to the "seek"
commands, which are also capable of switching songs.

Closes #119

Our previous use of Queue::SwapOrders() could cause surprising
results:

- sometimes, the old "current" song would be played again (if the
  newly selected song had not been played already)

- sometimes, the old "current" song would not be played again (if the
  newly selected song had already been played)

This is inconsistent, because it should not depend on whether the
newly selected song had already been played.

So instead of Queue::SwapOrders() we now use Queue::MoveOrderAfter()
and Queue::MoveOrderBefore(), which is more expensive, but also more
consistent.  It attempts to retain as much from the previous order
list as possible, and only moves the newly selected song around.

If an early exception gets caught (e.g. from
AllocatedPath::FromUTF8Throw()) before
DecoderControl::CommandFinishedLocked() is called, the decoder thread
would go in an endless loop, because DecoderCommand::START is still
set.

Closes #118

GME starts all track indexes at zero, but subtune prefixes
start at one. This fixes an off-by-one error during track
enumeration.

Our IcuCaseFold() fallback using strxfrm() is not actually case
insensitive.  This commit fixes the problem by switching to
strcasecmp().  That function is not guaranteed to support UTF-8, but
it's the best we can do in this sparse situation.

Closes #111

ffmpeg plugin: when decoded stream duration is unavailable, attempt fallback to container duration (fix MusicPlayerDaemon/MPD#110)

Closes #107

Workaround for the following LAME build failure:

 error: inlining failed in call to always_inline '_mm_sqrt_ps': target
 specific option mismatch

This is because the LAME build scripts do not check whether SSE is
available; they only check for the presence of the "xmmintrin.h"
header.

Requiring a Pentium 3 CPU is reasonable enough, and it's the first CPU
to feature SSE support.

This fixes an assertion failure caused by resuming playback before the
decoder has finished startup.

.. and not the stack buffer.  This made the AllocatedString destructor
crash.

Closes #52

Was broken by commit 8a86460b

Closes #87