[QingyuLiu0521]

This PR updates src/f5_tts/model/backbones/dit.py to preserve correct variable-length text conditioning while removing per-sample Python loops in text embedding.

• It switches to batched text_embed(text, seq_lens, ...), supports per-sample seq_len masking, and zeroes tail embeddings after lookup (p >= seq_len[i]).
• It applies positional encoding only on valid positions per sample and fixes average upsampling to use per-sample target lengths.
• Result: same intended behavior for mixed-length batches, with lower overhead and better training throughput.