Fix SpeedySpeech frontend and duration parity

Regenerated with the checkpoint-native Gruut English lexicon and corrected input padding for the duration predictor. Native output now matches the PyTorch reference waveform at 0.99999986 correlation.

Coqui-SpeedySpeech-LJSpeech-GGUF/coqui-speedy-speech-ljspeech-f32.gguf

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:14f1a4bf9b3169f492291f04e04f924155fae13ff30b0a91df3e4c8f58d137c8
-size 21682816
+oid sha256:7a5d9f566b282c4a8275b44637b887776d6c5baa29021c420d7707d41938cfb9
+size 24080352