Variables are marked with this annotation to make it clear that access
to the variable should be done through atomics.

I've also review the use of volatile in this commit.

Public issue https://github.com/mackron/miniaudio/issues/259

This was setting the previous pointer of newly attached nodes to NULL
instead of a pointer to the dummy head node. This then results in the
dummy head node never being updated when the node is detached which in
turn results in an uninitialized node being dereferenced.

This would try reading from an uninitialized buffer thereby resulting
in a bad audio glitch. This does two things to fix the problem:

  1) When there are no input nodes attached to an input bus, nothing is
     read and 0 will be returned for the frames read variable.

  2) The buffer is silenced by default.

This fixes a bug with ma_engine where it would glitch in the moment
just after the engine is initialized and before a sound or group is
attached.

This is no longer necessary and just introduces unnecessary latency and
inefficiency due to extra data movement.

This provides an optimization by allowing processing to bypass the
resampler. Audio data needs to pass through the resampler even for the
case where pitch=1 because it needs to update internal buffers which if
it didn't do, would result in a glitch when moving away from 1.

In practice most sounds won't require individual pitch control, however
in the interest in being consistent with miniaudio's philosophy of
things "Just Working", pitching is enabled by default. Pitching can be
disabled with MA_SOUND_FLAG_DISABLE_PITCH in ma_sound_init_*() and
ma_sound_group_init().

This is the first step towards decoupling the ma_effect API from the
engine for the eventual removal.

This also fixes bugs regarding channel conversion when processing an
engine node.

This bug happens when the channel count of the data source does not
match the channel count of the engine.

This is mainly for consistency with the node API, but also because it
more clearly indicates that it's an absolute time rather than a delay
which sounds more like a relative time.

This changes ma_sound_set_start/stop_delay() to take an absolute time
in frames based on the global clock. Previously these took a relative
time in milliseconds. To use a relative time, add it to the value
returned by ma_engine_get_time(). To use milliseconds, use a standard
sample rate to milliseconds conversion.

These changes are in preparation for fixing some issues relating to
retrieval of channel counts from data sources. The problem relates to
the asynchronous nature of the resource manager and how data sources
may be in the middle of loading when trying to initialize a sound which
results in the channel count not yet being available. The channel count
is necessary in order for the engine to be able to convert the data
source to the channel count of the final output.

This commit has known issues which will be fixed in the next commits.

This is needed for scheduling frame-exact starting and stopping of
nodes, in addition to any kind of time-based effects required by custom
nodes, such as fading.

This is not fully functional as of this commit.

The sample format is now always 32-bit floating point. This was done
for simplicity and practicality.

Effects will now be implemented by plugging effect nodes into the node
graph at the appropriate location.

With this change, attachment and detachment of a node's output bus is
now consistent:

  * ma_node_attach_output_bus()
  * ma_node_detach_output_bus()

The old function that allowed you to attach the other way around has
been removed as the caller can easily do that themselves by swapping
the order of parameters.

This changes the way ma_node_uninit() waits for the audio thread. Now,
instead of waiting for the *entire* graph to complete, it only waits
for it's local processing to complete.

This makes it possible to give individual input and output buses unique
channel counts which might be useful for specific kinds of effects.