MAINT: Clean adaptive algorithms

[stsievert]

What does this PR implement? It runs experiments to show how the Salmon server performs as the number of users varies. This PR simulates n_users crowdsourcing participants answering questions.

TODO:

• [x] Base implementation.
• [ ] Run with various configs (mostly vary n_users).
• [ ] Visualization & integration into docs.

This PR no longer does the above. Instead, it fixes some bugs with the adaptive implementation. In addition, it does the following:

• [x] API: adds endpoint for alg timings
• [x] BUG: fixes error with epsilon.
• [x] MAINT: flips the queries randomly.
• [x] API: allows targets to be an integer N to create N targets
• [x] VIZ: Visualizes algorithm timing on the dashboard.
• [x] DOC: Adds dashboard to documentation.

Reference issues/PRs #65 is necessary for this PR. That PR verifies that adaptive gains are present offline when the time required for model updates aren't considered. This PR isn't nearly as deterministic.

I need to rerun #65 to ensure that those adaptive gains are realizable.

[stsievert]

It's not uncommon to see about 5 queries with the same head appear in a row in this implementation. That requires more investigation. Are the query scores negated? Are the searches too complete?

I do recover a 1D manifold with the strange fruit example. After ~700 examples, 10 elements on the end of the manifold are [9, 8, 7, 6, 5, 3, 4, 0, 10] from distal to proximal. I think I'd expect that order to be reversed. I've also noticed that with a moderate number of responses (~700) the embedding conforms very nicely with a straight line. With more responses (~1800) that's less true; the embedding is not as much of a manifold; there's more variation.

I think the best way to determine if the scores are negated is to measure the accuracy of the embedding throughout time for both implementations (negated and non-negated scores).

[stsievert]

One example for the negated scores: I saw a query with head 36 and left/right of 12/13 with this embedding:

This embedding is after 3320 responses (long after the 1D manifold was present). It seems this query is meaningless: the distance from 36 and 12 is nearly the same as the distance from 36 and 13.