fix: add missing parentheses to @torch.inference_mode decorator

[jnMetaCode]

In ChatTTS/model/tokenizer.py, the decode method uses @torch.inference_mode without parentheses, while the encode method (and every other usage across the codebase) correctly uses @torch.inference_mode().

Without parentheses, torch.inference_mode is applied as a raw class rather than as a decorator factory. This means the decode method gets replaced by the torch.inference_mode context manager object itself, instead of being wrapped with inference mode enabled. As a result, calling decode won't actually perform any decoding — it silently returns the context manager.

This patch adds the missing () to match the rest of the codebase and restore correct behavior.