fix: allow oversized Vulkan parameter tensors

[leejet]

Summary

• Treat backend max buffer size as a chunking threshold instead of a hard per-tensor limit when allocating parameter buffers.
• Allow oversized tensors to be allocated alone so Vulkan can handle large weights such as Qwen embedding tensors.

Related Issue / Discussion

Fix #1659.

Additional Information

N/A

Checklist

• I have read and confirmed this PR follows the contribution guidelines.