Add more JOIN strategies

[Jackie-Jiang]

Currently we support 4 JOIN algorithms:

1. Hash: hash partition both sides and shuffle to the same worker to join
2. Random left + Broadcast right: for non-equal join, random distribute left side and broadcast right side to join
3. Colocated: when left and right table are partitioned the same way, join locally without shuffling
4. Lookup: when right table is a dimension table, and join is equality join on right table's primary key, join locally by looking up the lookup table of the dimension table

Both 3 and 4 can avoid data shuffling, and we want to add more strategies that can reduce data shuffling:

1. When right table is fully replicated to all servers (similar to dimension table, but not necessary with a lookup table), we can join locally
2. Broadcast the right table, and join on left table local worker
3. When left table is partitioned and the join is on the partition key, partition the right table, and join on left table local worker

[Jackie-Jiang]

cc @gortiz @yashmayya @bziobrowski