ENH: Add sparse middle-out reconciliation via ``MiddleOutSparse``

[christophertitchen]

This PR adds a middle-out reconciliation class called MiddleOutSparse for sparse "summing" matrices, $S$, to significantly reduce the wall time for reconciling large hierarchical structures from a "middle" level, which now take at least $99.6 \%$ less time to reconcile.

• Uses the data structure encoded in $S$ to efficiently identify the aggregation pathways in the subgraph terminating at the "middle" level to create a sparse $S$ of the subgraph.
• Performs bottom-up reconciliation via BottomUpSparse and allocates the reconciled point forecasts in a memory-efficient manner.
• Uses the data structure encoded in $S$ to efficiently identify the subgraphs for each cut node in the "middle" level by identifying their respective leaf nodes and slicing $S$ to find the remaining nodes, which are then converted to a CSR matrices.
• Constructs the tags argument only if required for out-of-sample disaggregation methods, i.e. forecast proportions.
• Avoids repeated checks for the strictness of the hierarchical data structure by setting tags to None for in-sample methods.
• Finds the "horizontal" intersection of the nodes for each level and the subgraph for each cut node to efficiently construct the tags argument for the subgraph in the out-of-sample case.
• Performs top-down reconciliation via TopDownSparse for each cut node and allocates the reconciled point forecasts in a memory-efficient manner.
• Tests the output of MiddleOutSparse and equivalence with MiddleOut up to the machine epsilon of a double.

Dataset	$\textup{dim} \left ( P \right )$	$K$	$T$	$h$	$\ell$	Method	MiddleOut	MiddleOutSparse	Reduction
M5	$\small{\left( 30\, 490 \times 42\, 840 \right)}$	$12$	$1\, 969$	$28$	$9$	A #277	$\gt 2\, hr$ *	$28.1\, s$	$\gt \, 99.6 \%$ *

* This was cancelled after 2 hours. 😴

[review-notebook-app[bot]]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB

[christophertitchen]

Thanks, again great work! Looking forward to seeing what you're up to next :)

I think after all your work it would be awesome if we update the docs and have a tutorial showing the nice sparse methods that you implemented. Let me know if you'd like to work on that, otherwise I'll pick that up sometime.

Thank you for your reviews and great suggestions, Olivier!

I would be happy to both update the docs and create a tutorial for the sparse methods.

[elephaint]

Thanks, again great work! Looking forward to seeing what you're up to next :) I think after all your work it would be awesome if we update the docs and have a tutorial showing the nice sparse methods that you implemented. Let me know if you'd like to work on that, otherwise I'll pick that up sometime.

Thank you for your reviews and great suggestions, Olivier!

I would be happy to both update the docs and create a tutorial for the sparse methods.

Great! Let's do that in another PR, I'll merge this one then.