Enable Flink buffer debloating to reduce checkpoint alignment time under backpressure

Steps

1. Set taskmanager.network.memory.buffer-debloat.enabled: true — it defaults to false and must be explicitly turned on.
2. Set taskmanager.network.memory.buffer-debloat.target to the desired time-to-consume in-flight buffered data, rather than relying on a fixed byte-size buffer.
3. Tune sampling behavior with buffer-debloat.period and buffer-debloat.samples if the default cadence doesn't react fast enough to changing throughput.
4. Adjust buffer-debloat.threshold-percentages if buffer size oscillates too aggressively between recalculations.
5. Expect the biggest win on aligned checkpoints, since debloating shrinks the volume of in-flight data that has to be aligned; it still works with unaligned checkpoints but the benefit is smaller there.