* Data is not deleted after expiration due to connected readers

*Problem*

A problem is observed when stress testing pulsar using [pulsar-flink](https://github.com/streamnative/pulsar-flink) -
No matter what TTL or retention setting is used, the data is never cleaned up. So the stress test ends up failing due
to disk filled up.

The root cause of the problem is described as below.

when a reader is opened using `MessageId.earliest`, a non-durable cursor with position (-1, -2) is added to the cursor heap.
The position `(-1, -2)` in the heap is never updated because non-durable cursors are never advanced when mark-deletions
happen. So the slowest cursor position is always `(-1, -2)`, thus causing no ledger can be deleted even they are expired
or over quota.

*Motivation*

Fix the problem to make sure Pulsar honor to TTL and retention settings.

*Modifications*

- Fix the `startPosition` when PersistentTopic opens a non-durable cursor on `MessageId.earliest`.
  So the `startPosition` is (-1, -1) not (-1, -2).

- Fix the `NonDurableCursorImpl` constructor to check if the position in the ledger of `MessageId.earliest`.
  If the provided position is in the `earliest` ledger, the mark-deleted position will be set to the previous
  position of first position.

- Fix the `NonDurableCursorImpl` to advance ledger cursor when mark-deletion happens on a non-durable cursor.

*Verify this change*

Unit tests are coming.

(cherry picked from commit 3e7cb68b)