Make combine_similar less greedy for merge

[phofl]

This makes combine_similar less greedy, we are currently removing projections on both sides and see if we are equal, we should do one after another to avoid pushing projections up unnecessarily. Shuffling is expensive, so this should help quite a bit.

This sits on top of #333, the previous structure made this change even harder that it was anyway

[rjzamora]

Is there a reproducer for the behavior being targeted by this PR? This test seems to pass on main for me.

[phofl]

Yeah the bug was blocked by the additional layer in #333 but shows up now

[phofl]

Previous tree with the reproducer:

Assign: key='new'
  Projection: columns=['b', 'x', 'y']
    Projection: columns=['b', 'x', 'd', 'y']
      HashJoinP2P: left_on='b' right_on='d' left_index=False right_index=False shuffle_left_on='b' shuffle_right_on='d'
        HashJoinP2P: left_on=['a'] right_on=['a'] left_index=False right_index=False shuffle_left_on=['a'] shuffle_right_on=['a']
          FromPandas: frame='<pandas>' npartitions=2 columns=['a', 'b']
          FromPandas: frame='<pandas>' npartitions=3
        FromPandas: frame='<pandas>' npartitions=3 columns=['d', 'y']
  Add:
    Projection: columns='b'
      Projection: columns=['b', 'a', 'd']
        HashJoinP2P: left_on='b' right_on='d' left_index=False right_index=False shuffle_left_on='b' shuffle_right_on='d'
          HashJoinP2P: left_on=['a'] right_on=['a'] left_index=False right_index=False shuffle_left_on=['a'] shuffle_right_on=['a']
            FromPandas: frame='<pandas>' npartitions=2 columns=['a', 'b']
            FromPandas: frame='<pandas>' npartitions=3
          FromPandas: frame='<pandas>' npartitions=3 columns=['d', 'y']
    Projection: columns='x'
      Projection: columns=['b', 'x', 'd']
        HashJoinP2P: left_on='b' right_on='d' left_index=False right_index=False shuffle_left_on='b' shuffle_right_on='d'
          HashJoinP2P: left_on=['a'] right_on=['a'] left_index=False right_index=False shuffle_left_on=['a'] shuffle_right_on=['a']
            FromPandas: frame='<pandas>' npartitions=2 columns=['a', 'b']
            FromPandas: frame='<pandas>' npartitions=3
          FromPandas: frame='<pandas>' npartitions=3 columns=['d', 'y']

We have the 2 Join layers right next to each other without dropping unnecessary columns

The new tree:

Assign: key='new'
  Projection: columns=['b', 'x', 'y']
    HashJoinP2P: left_on='b' right_on='d' left_index=False right_index=False shuffle_left_on='b' shuffle_right_on='d'
      Projection: columns=['b', 'x']
        HashJoinP2P: left_on=['a'] right_on=['a'] left_index=False right_index=False shuffle_left_on=['a'] shuffle_right_on=['a']
          FromPandas: frame='<pandas>' npartitions=2 columns=['a', 'b']
          FromPandas: frame='<pandas>' npartitions=3
      FromPandas: frame='<pandas>' npartitions=3 columns=['d', 'y']
  Add:
    Projection: columns='b'
      HashJoinP2P: left_on='b' right_on='d' left_index=False right_index=False shuffle_left_on='b' shuffle_right_on='d'
        Projection: columns=['b']
          HashJoinP2P: left_on=['a'] right_on=['a'] left_index=False right_index=False shuffle_left_on=['a'] shuffle_right_on=['a']
            FromPandas: frame='<pandas>' npartitions=2 columns=['a', 'b']
            Projection: columns=['a']
              FromPandas: frame='<pandas>' npartitions=3
        Projection: columns=['d']
          FromPandas: frame='<pandas>' npartitions=3 columns=['d', 'y']
    Projection: columns='x'
      HashJoinP2P: left_on='b' right_on='d' left_index=False right_index=False shuffle_left_on='b' shuffle_right_on='d'
        Projection: columns=['b', 'x']
          HashJoinP2P: left_on=['a'] right_on=['a'] left_index=False right_index=False shuffle_left_on=['a'] shuffle_right_on=['a']
            FromPandas: frame='<pandas>' npartitions=2 columns=['a', 'b']
            FromPandas: frame='<pandas>' npartitions=3
        Projection: columns=['d']
          FromPandas: frame='<pandas>' npartitions=3 columns=['d', 'y']

We have an intermediate Projection that drops unnecessary columns right away

[rjzamora]

This logic is pretty hard to follow (mostly because it was already complicated), but I don't have any better ideas, and I don't think it is worth the effort to spend much time attempting a simplification until we have addressed #336

[rjzamora]

Nit: Rename to _PARTITION_COLUMN for clarity.

[phofl]

Oh good point, thx