Keep Tau error cleanup within its original deadline on cancellation

[bradhilton]

A second parent cancellation could abandon an environment deletion already cleaning up an earlier rollout failure. Continue waiting for cooperative deletion within the original absolute 30-second deadline, then preserve the primary error and later cancellation together. Repeated cancellation does not renew the deadline.

The existing deadline fallback is retained: an unfinished deletion is cancelled and its result consumed by the existing callback. This does not promise completion of cancellation-resistant deletion or change success-path deletion.

Validation: all 39 Tau client tests pass. Five new cases fail on the prior source and cover ordinary/cancellation primary errors, repeated cancellation, cleanup failure, and the decreasing timeout budget. Existing deadline and resistant-delete tests remain unchanged. Ruff, formatting and affected type checks pass; independent source review is clear. Fake client only, with no provider calls or live-source changes.

Downstream validation: nine offline actual-method controls cover maintained Caladan and the pending evaluation helper in OpenPipe/caladan#844. Ordinary induced cancellation remains an isolated incomplete evaluation; a mixed body error and later cancellation preserves the prior fatal outcome while completing DELETE. No dependency on the separate dynamics cleanup PR was established. Frozen experiment snapshots require their own adoption review.