Refactor TaskError to hold Control and NonControl

[Aaron1011]

This will allow us to reduce error type duplication in tensorzero/tensorzero by removing the NonControlToolError type

---

Note

Medium Risk
Medium risk because it changes the core error type used across task execution and worker failure handling, including retryability and failure serialization semantics, which may be a breaking API change for downstream users.

Overview
Refactors task execution errors so TaskError becomes a wrapper enum with Control(ControlFlow) vs NonControl(NonControlTaskError), moving all failure variants (timeout, validation, step failure, child failure/cancel, DB/serialization, etc.) into NonControlTaskError with centralized retryable() logic.

Updates task/worker/heartbeat code to construct and propagate TaskError::NonControl(...) for real failures while preserving ControlFlow behavior for suspend/cancel/lease-expired, and adjusts error serialization by delegating failure JSON encoding to serialize_non_control_task_error. Also re-exports NonControlTaskError from lib.rs.

^{Written by Cursor Bugbot for commit 5e773a6. This will update automatically on new commits. Configure here.}