Consider query-only test time training

**Is your feature request related to a problem? Please describe.**
Look at Bansal etal, ICLR 2026: "Test-time training for long-context LLMs". How difficult is this to implement?

Start by reading the cited Bansal et al. ICLR 2026 paper and then inspect the repository's existing training and inference entry points. Clarify what query-only test-time training should change, its required inputs and evaluation criteria; the work is done when the scope and implementation approach are agreed.

Written by the indexing model from the issue text.