fix: gracefully fall back flash_attention_2 -> sdpa when flash_attn not installed

fix: default attn to sdpa (no flash_attn dependency at inference)

docs: works on modern transformers now; soften version note

fix: cross-version transformers compat (resolves list|set crash on modern transformers)

docs: pin transformers==4.57.1 + fix usage (resolves list|set / missing lm_head on newer transformers)

roxi-tts-v2: MOSS-TTS-Nano LoRA Indian-English (with IIT-M attribution)

initial commit