Remember best known statistics for TableHandle on engine side

[findepi]

As described in #844, predicate pushdown may degrade the quality of statistics provided by a connector.
If the connector is not able to consider Filter in the returned response, and engine no longer has a Filter, the statistics will not reflect the effect of filtering.

This can become a bigger issue, as we progress with #18 and #4249, #6613 #6620 in particular.

Here, a solution is proposed where Engine keeps track of pre-pushdown statistics on TableHandle / TableScanNode and uses them when connector can no longer produce statistics.
By "pre-pushdown" i mean statistics derived from the TableScan + the operation being pushed down (e.g. Filter, Aggregation, Join, TopN, etc.), calculated on the engine side.

cc @rzeyde-varada @losipiuk @sopel39 @martint @kokosing

[findepi]

The computation would need to happen (eagerly, or lazily - to be defined), after successful pushdown, based on the plan nodes before the pushdown.

[sopel39]

This does not exactly solve #844 as connector might support predicate pushdown but might not support predicate pushdown for stats computation.
For #844 I think engine should first try to compute stats with predicate applied and if it returns empty, then engine should try to compute stats without any predicate and apply predicate with stats calculator.

Anyways, I don't think io.trino.spi.connector.ConnectorMetadata#getTableStatistics is used (and implemented) correctly (e.g see io.trino.plugin.hive.HiveMetadata#getTableStatistics)