Accelerate module cache copying, support RoPE encoding and Qwen3 LLM

[Panxy777]

This PR improves the efficiency and compatibility of PromptCache.update by reducing redundant KV copy operations, minimizing CPU–GPU transfer overhead, and adding support for Qwen3 series models and RoPE encoding.

Key changes:

• Retain only the common prefix with the previously staged cache; re-copy KV for divergent suffix segments.
• Prefetch module caches to the target GPU before reuse, avoiding repeated CPU-to-GPU transfers.
• Replace per-module copy with per-layer batched copy:
  • Upload all updated modules to the target device in advance.
  • Concatenate KV tensors across modules along the sequence dimension per layer.
  • Perform a single write per layer into the destination cache slice.
• Add a fast path for single-module updates to avoid unnecessary concatenation.
• Extend support to Qwen3 models and RoPE encoding, ensuring compatibility with their KV structures.

Benefits:

• Reduce copy operations from O(#modules × #layers) to ~O(#layers).
• Lower kernel launch overhead and Python loop cost.
• Significantly improve performance in multi-segment suffix updates (e.g., tag switching scenarios).
• Broader model support including Qwen3 and RoPE-based architectures.