0.45.0

• GPU narration on Windows for every graphics card, in a 343 MB download
  instead of 2 GB. Windows now uses DirectML, which runs on AMD, Intel and
  NVIDIA cards alike, in place of CUDA; Linux keeps CUDA for NVIDIA cards. The
  pack brings a copy of Kokoro made to run on DirectML, with the same voice.
  On a GTX 1060, 0.12 of real time once warm, with the processor nearly idle.
• The fastest card is found by trying. DirectML takes the first adapter
  Windows lists, which on most laptops is the integrated chip, and there
  Kokoro can run twelve times slower than real time. The first time a voice
  is used, each card is tried in a separate copy of the app and the fastest
  one that keeps ahead of the voice is kept; if none does, the processor
  narrates and Settings says so.
• A card that fails can't take the app down. Some graphics failures end
  the whole process instead of reporting an error, so a voice is only ever
  narrated on a card after that separate copy has spoken with it.
• GPU files from 0.44.0 can be removed from the same place in Settings.
• Settings are written far less often. Every change rewrote its whole
  file, fifty times for one drag of a slider; now a save waits for the ones
  already queued.