Speedup thicket composition by performing profile unions in binary tree order

[michaelmckinsey1]

This PR extends off #169. Limiting the total size of the ensemble during unification is more impactful than the performance gains from parallelism, and this solution is less complicated than using multiprocessing.

	Parallel Sorting	LBANN
Profiles	15734	16
Input Nodes (min, max)	5, 18	6266, 6889
Input Dataframe Size (min, max)	''	''
Output Nodes	154	7106
Output DF	2423036	113696

Dataset	Thicket Version	Time	Unions	old_to_new size (min, max)
Parallel Sorting	Develop	>120m	15733	157459
	#170	12m	''	16, 185
	#185	33m	''	157459
LBANN	Develop	17m	15	106897
	#170	7m	''	12824, 13821
	#185	3m	''	106897

Number of unions is not affected by the different unifying strategy in this PR. This is because:

# PR170 - 3 unions
tk1 U tk2 = tk12
tk3 U tk4 = tk34
tk12 U tk34 = tk1234

is the same as

# develop - 3 unions
tk1 U tk2 = tk12
tk12 U tk3 = tk123
tk123 U tk4 = tk1234

This PR increases performance by keeping the old_to_new mapping dictionary small as we call concat_thickets n-1 times instead of one time (as in the develop branch). The old_to_new dictionary requires an operation to update the dictionary mappings per each union, which can become expensive if the dictionary is too large. The old_to_new dictionary is necessary to replace the node objects in the performance data with the new node objects in the graph that results from union.

[ilumsden]

The code is good, but something has to be done to make the tree-based concatenation clearer. At bare minimum, this needs to be well commented, otherwise no one is going to have a clue what this is doing.

[ilumsden]

Using tqdm here makes this process very hard to understand. I'd suggest not using tqdm and/or extensively documenting this with comments.

[michaelmckinsey1]

tqdm is really useful to see what is happening for longer reads.

I changed it to use a while loop, which should be clearer with the popping and appending that is happening in the loop. And I added very detailed comments.

[ilumsden]

LGTM. The stuff you're doing with tqdm also makes much more sense now that you've added the comments and switched to a while loop

[ilumsden]

@pearce8 this PR is also ready for your review