Add Archa TTS (Qwen2.5-0.5B + SNAC) support

[Copilot]

Integrates YangNobody12/Archa-TTS-0.5B-th — a Thai TTS model built on Qwen2.5-0.5B with LoRA fine-tuning and SNAC 24kHz audio codec — as a new pretrained="archa" option.

New file: pythaitts/pretrained/archa_tts.py

• Loads Pakorn2112/Archa-TTS-0.5B-th and hubertsiuzdak/snac_24khz from HuggingFace
• Generates audio tokens autoregressively, decodes 7 tokens per SNAC frame into 24kHz waveform
• Auto-selects cuda/cpu; bfloat16 on supported GPUs, float32 on CPU
• Optional noise reduction via noisereduce (degrades gracefully if absent)
• Supports return_type="file" (WAV) and return_type="waveform" (numpy array)

Updated: pythaitts/__init__.py

• Adds archa branch in load_pretrained and tts() dispatch
• Passes device through so callers can select GPU/CPU

Usage

from pythaitts import TTS

tts = TTS(pretrained="archa", device="cuda")
tts.tts("สวัสดีครับ วันนี้อากาศดีมากเลย", filename="output.wav")

# or return numpy waveform
waveform = tts.tts("สวัสดี", return_type="waveform")

Required extras: pip install torch transformers snac soundfile noisereduce scipy numpy