Features

• server: New faster GPTNeoX implementation based on flash attention

Fix

• server: fix input-length discrepancy between Rust and Python tokenizers