Ensure correct handling of errors from thread-local storage

BoringCrypto exposes `ERR_clear_error()` which clears the thread-local error stack.

For any function in BoringCrypto which may set error state, we **must** read the error stack, before we do any async/await, as this may trigger the code to continue running on a different thread.

Perhaps, we have to use `_Scope.sync` for everything, and ensure that we check/clear the error stack at the end of every `_Scope.sync` block.
Or maybe we have to invent a new block form.