[fix] fix workspace mismatch when using fa2

[xenshinu]

No description provided.

[Esthesia]

Thank you for the update!

[matteso1]

this no longer builds against its own binding — graph.py:160 passes capture_begin(capture_error_mode=...) and the current foundry.ops binding requires a positional pool, so the branch needs a rebase before the fix can be evaluated. worth prioritizing: the issue #2 bug it fixes still reproduces byte-identically on sm_89 (details in #2 (comment) — that issue is closed but the bug is live). happy to validate a rebased version same-day on an RTX 4090 / driver 595.71.05 / CUDA 13 setup.