Tolk seems to fail on strings larger than 325 characters in length. Here we:
* Send any strings with 300 or fewer characters through directly.
* For larger strings, split on whitespace boundaries, then create and send buffers of 300 or fewer characters.
This may not handle internationalized text, and may not handle someone bombarding TTS with a giant word. PRs for either welcome.
Two Tolk instances were being created. One checked for the presence of a screen reader. The other actually performed the speech, and was returned as part of the `TTS` instance.
Unfortunately, Tolk doesn't seem to appreciate being called twice. So here we check if a screen reader is detected and, if one is, return the instance that did the detection. Otherwise, error out and return the WinRT backend.
I'd previously attempted to normalize everything to `u8`, but this had some drawbacks:
* It failed to account for some synthesis drivers defining normal as mid-range, while most define it very low.
* It didn't track the normal value for a given synthesizer.
* There was no clean way to map a curve between the minimum, normal, and maximum rates.
Here we track the minimum, normal, and maximum values of rate, pitch, and volume. Sanity checks are done on set.
Also, as a further proof-of-concept, all properties are now implemented for the WinRT driver.