EndOfInput was used to handle any stopping of the job. When
stopping with savepoint the input is not actually ended.
This causes issues with some sinks (e.g. Iceberg).

With this change, endInput is not call for stop-with-savepoint.

To differentiate stop-with-savepoint from other cases
only checkpoint (RPC/barriers) are considered and not network EOP.
That's enough because EOP is only injected after the CP completion
(i.e. when the downstream is also notified by sync savepoint by CP
barriers).

…h chaining strategy

…ent EndOfInput

…ally