fix(executor): track For Each iteration bodies in pendingTasks

[eskp]

Summary

• Stop parallel For Each iterations from silently dropping body steps when the workflow SDK's checkpoint resume severs an iteration's await chain after its first body step. The downstream step never got scheduled, the iteration still reported success, and the run finalised with status success missing the dropped work.
• Each iteration body's executeBodyNode call is now wrapped in pendingTasks.track, so the workflow-end drain holds a strong reference to in-flight body chains and recovers orphaned continuations. Sequential mode is unaffected. Mirrors the KEEP-395 fix that protects the main DAG from the same class of SDK truncation.

Test plan

• pnpm vitest run tests/unit/executor-drain-loop.test.ts -- new simulates parallel For Each iterations: drain rescues truncated body recursion case passes alongside the existing KEEP-395 drain-loop tests
• pnpm vitest run tests/unit/for-each-body-runner.test.ts tests/unit/for-each-body-recursion.test.ts tests/unit/for-each-executor.test.ts tests/unit/for-each-concurrency.test.ts -- no regressions
• pnpm type-check -- clean
• Run the user-reported test-parallel-iteration-bug workflow on a deployed branch and confirm parallel mode records all body steps (e.g., 4 iterations -> 4 decode + 4 HTTP) instead of dropping iterations after the first step