CORE-45: inline dependency resolution + eager tasks for async graph execution

[LawlAoux]

Summary

• Nodes whose dependencies are already computed now execute inline instead of each being wrapped in asyncio.create_task (~24k tasks/turn on large voice bots, none of which start executing until the topological loop ends); a node blocked on a genuinely pending dependency falls back to a task created with Python 3.14's eager_start=True, so it runs at creation and network calls fire mid-traversal.
• Input-option priority, pruning, and exception semantics are preserved: node_to_input_inline mirrors node_to_input_async, returning _PENDING to route to the task path. get_deps_and_await is dead in the new model and removed.
• bon_secours_voice (45,540 nodes): CG framework overhead per turn 493–504ms → 364ms (−26%); tasks/turn 24,183 → ~100 on turns without real I/O wait; wall time on network-free turns −9–12% (connect 1305→1183ms). Bot responses byte-identical across old and new runners.

Rebase onto master 78 (22 Sep)

• Sync-chain fusion (feat: fusing chains of sync nodes that are downstream to async nodes #85) was reverted in Revert "Merge pull request #85 from hyroai/yinon/poc-fuse-chains" #98, so the fusion notes this description used to carry, and the two fusion tests this PR had adapted, are gone with it. run.py merged cleanly; version 79, uv.lock re-locked.
• Re-benchmarked with nlu-runtime's network-free per-turn harness (model endpoints and LLM calls answered in-process, one bot per process, N=7, this branch vs the installed CG 78 back-to-back on the same nlu code; laptop numbers, so read the percentages): connect p50 −21% / −30% / −27% on atlantic / louisiana / bon_secours, greet −12% / −23% / −13%, turns with an embed + SLM call −6…−11%. Turns whose routing-classifier call suspends mid-graph are unchanged (±1%): their downstream falls back to the task path, as designed. The server's per-message gc.collect(0) drops from 12–37ms to under 1ms on connect/greet turns and to ~3ms on embed turns, since 24k fewer Task objects are created per turn. Bot responses byte-identical on all three bots, zero errors.
• Gates on the rebased branch: CG suite 119 passed; nlu-runtime slot_filling + listen suites 587 passed against it.

Jira

https://hyro-ai.atlassian.net/browse/CORE-45