chore: bump interp-engine to 1.7.3 in inference, graph and nla

[hijohnnylin]

Pins `interp-engine` 1.7.0 → 1.7.3 in the three Python apps (pyproject + uv.lock, relocked with `uv lock --upgrade-package interp-engine`, no other resolution changes) and updates the version noted in AGENTS.md.

What 1.7.3 brings (decoderesearch/interp-engine):

• `Tokenize.message_spans` tags a system turn the chat template injects on its own (Llama's knowledge-cutoff preamble, Qwen2.5's default persona) as `role=system` with `message_index=None`, so the lens and NLA chats can show it as its own bubble instead of folding it into the first user message's header. `message_partition` is unchanged, so pooled activations are unaffected.
• 1.7.1–1.7.2: capture-side fixes (token ids moved to the model device before the forward; a vLLM static-checkpoint refusal).

Rollout: the webapp side (system bubble, pairing bubbles to messages by `message_index`) is already on main and works with both the old and the new engine, so the inference and NLA servers can roll onto this at any pace. Cached jlens/NLA results keep their stored spans and render as before.

Made with Cursor