Race condition in HarmonyEncoding concurrent access

[xiaogecai]

Problem

Multiple threads concurrently access HarmonyEncoding tokenizer file downloads, causing race conditions.

Root Cause

No mutex-based file locking during concurrent tokenizer downloads.

Expected Solution

• Add thread-safe initialization with OnceLock
• Use DOWNLOAD_MUTEX for file locking
• Support offline loading API (refs Race condition in HarmonyEncoding concurrent access #1)

Refs: #6, #1

[xiaogecai]

PR #2 implements std::sync::Mutex with OnceLock for thread-safe initialization. The DOWNLOAD_MUTEX prevents concurrent download races. Offline loading from issue #1 is included.