This commit includes the dInitScan cost for every rebind in a bitmap table
scan.

This is a partial revert of commit : 1cae6057

It turns out including the init scan cost for every rebind more accurately
represents executor behavior. Not including the init scan cost per rebind led
to picking up much slower Index Nested Loop Join over Hash Join.

In the example minidump included in the commit, the Index Nested Loop Join runs
~ 3 times slower than the Hash Join, yet it was getting picked. In larger
tables, sometimes we ended up picking Index Nested Loop Join which never
finishes executing.
Co-authored-by: NSambitesh Dash <sdash@pivotal.io>