This adds two dependencies, they are only needed for the test suite.
Compile time will not increase for users. The extra overhead when
running the test suite is worth it imho. The test file is significantly
shorter and there is less code duplication. A run time solution would
decrease the test interface (you would have to manually find out which
params caused the test).
As I understand it, this is the issue:
* Previously, when a queue has `keep_alive_if_empty` set to true, and it becomes empty, then it will push a silence lasting 10ms onto the queue.
* This is an issue because `current_frame_len` would have returned the worst case, `512`, and the silence lasts less than that.
* This means that unless the source is added immediately to the queue, and so a silence is never played, then the first actual source could start playing at a frame that is not aligned to its channels, or play at the wrong sample rate.
* This is only determined by when the source is added to the queue after its initialization. This explains why the issue was inconsistent, as it relied on the speed of execution of code which is basically random.
Solution
* Change the functionality of `Zero` to add a method to create a silence with a certain number of frames.
* Replace the 10ms silence with a silence the length of `THRESHOLD`
* Change queue's `current_frame_len` to return `THRESHOLD` if a silence will be played next.
