Fix Kokoro ZH model shape mismatch and add mixed language support

[Alex-Wengg]

Summary

• Fix check_array_shape to properly detect MLX vs PyTorch conv weight formats for 1D convolutions
• Update weight sanitization in kokoro.py and istftnet.py to use format detection instead of unconditional transpose
• Add Chinese-to-Bopomofo conversion using pypinyin for ZH model compatibility (the ZH model uses Bopomofo symbols, not IPA)
• Add number-to-Chinese conversion for proper TTS of numeric content (e.g., "23" → "二十三")
• Add mixed Chinese/English text processing in pipeline
• Update tests for check_array_shape function

Test plan

• Verify Kokoro-82M-v1.1-zh model loads without shape mismatch errors
• Test Chinese TTS output with lang_code="z"
• Test mixed Chinese/English text (e.g., "今天天气很好。Hello, how are you?")
• Test numbers in Chinese text (e.g., "大概是23度左右")
• Verify transcription of generated audio matches input