Use ShardFilter to reduce calls on DescribeStream

Is your feature request related to a problem? Please describe.

The dynamodb source must discover new shards. It currently does this every 1 minute (though this is now configurable per #7000).

The LeaderScheduler runs a loop running on the lease interval. It calls DescribeStream to get all the shards.

There are two problems with this approach:

1. This leads to latency spikes. When a ShardConsumer reaches the end of a shard its children are not discovered until the next discovery cycle. When this happens, there can be a pause in ingestion.
2. Using the DescribeStream to list all shards can be inefficient and lead to throttling.

Problem 1 can be partially mitigated by #7000 which will reduce the latency. But the latency still exists - it is just shortened. And using this can possibly exacerbate problem 2 by calling DescribeStream too often.

Describe the solution you'd like

DynamoDB's DescribeStream provides a ShardFilter.

We can change the design such that when a shard consumer reaches the end of a shard, it calls DescribeStream filtering by the shard's ID to fetch child shards. When it discovers these it can create the stream partitions immediately. This would give us child shards immediately rather than waiting for the full-stream DescribeStream.

If for some reason the ShardFilter call doesn't include a child shard (this can happen rarely), then the periodic discovery would still run and pick it up.

Describe alternatives you've considered (Optional)

We can rely on lowering the leasing interval as given in #7000. But we risk throttling.

Additional context

The DynamoDBStreamsDataFetcher takes a similar approach:

https://github.com/awslabs/dynamodb-streams-kinesis-adapter/blob/master/src/main/java/com/amazonaws/services/dynamodbv2/streamsadapter/DynamoDBStreamsDataFetcher.java#L367