Restrictions on soundio_{in,out}stream_get_latency()

The documentation for the two `soundio_{in,out}stream_get_latency()` functions specifies that they must only be called from the read or write callbacks, respectively. (Presumably due to the fact that they access non-thread-safe variables maintained by the capture/playback thread.)

I think this is prone to misuse---many developers will no doubt call this function from their application thread, and when told that this is incorrect, respond with "Hey, it works fine... it's never given me a problem."

I'd like to suggest the possibility of adding a `double *out_latency` parameter to the read/write begin or end functions. The user can pass in a variable pointer if they want to know the latency, or NULL otherwise. Because the means of obtaining this information is one and the same with the begin/end functions, there is no separate function to misuse.

Also: The documentation for `soundio_outstream_get_latency()` indicates that it is normally called right after `soundio_outstream_end_write()`, suggesting that this function should be the one to gain the `out_latency` parameter. But perhaps it may be more useful if the latency can be obtained from `_begin_write()`, because then the write callback would be able to conditionally write one set of samples or another depending on the specific value of that latency.