fix: handle GPU memory reports and LLM encoding failures

[leejet]

Summary

• Restrict the free-memory underflow rejection to Vulkan. Other backends remain capped by total memory minus tracked residency.
• Log reported memory and tracked allocations when capacity checks fail, and update the backend documentation.

Related Issue / Discussion

Try to fix #1989.

Additional Information

N/A

Checklist

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