API: implement SRR and various maintenance/improvement tasks.

[stsievert]

What does this PR implement? It allows the following:

• Implements an "synchronous round robin" scheme that searches n_search queries (instead of asynchronously).
• Specifying the triplets indices in validation sampling.
  • ...by specifying queries: [[0, 1, 2], [3, 4, 5]] in init.yaml under the samplers heading.
• Customizing the participant ID
  • ...by visiting http://localhost:8421/?puid=foobar.
• Allow customizing samplers_per_user
  • Related: API: allow /query?ident=.
  • ...by specifying samplers_per_user in init.yaml.
• Allow user to initialize embedding with pre-computed embedding
  • (which is especially relevant given the downloadable online embedding)
• Doc changes:
  • Clarify sampling key in init.yml.
  • Change "ARR" to mean "async RR" instead of "adaptive RR"
  • Add initial embedding to offline embedding
• Maint:
  • better error handling on dashboards/getting embeddings.
  • Better HTML serving of targets (prefetch instead of custom solution)
  • Adds torch.no_grad in score.

Backwards incompatible changes

• Renames RandomSampling to Random to fit with naming scheme of other samplers.
  • (but warns if wrong file uploaded)
• (internal) renames Runner to Sampler.

Reference issues/PRs This closes #114 and #117.

[stsievert]

@crcox does this resolve your issue? I think this is similar to the external ID/ext_id you mentioned in #117.

[crcox]

Yes, that will do the trick. Clarification on validation queries: is it something like

samplers: {
    Validation: {},
    queries: [[0, 1, 2], [3, 4, 5]]
}

or

samplers: {
    Validation: {
        queries: [[0, 1, 2], [3, 4, 5]]
    }
}

?

[stsievert]

basic.yaml shows an example of how to specify the indices. Here's a snippet:

samplers:
  ARR: {}
  Random: {}
  Validation:
    queries: [[0, 1, 2], [3, 4, 5], [6, 7, 8]]

(updated to reflect sampler renaming)

[crcox]

Perfect, thanks!

[stsievert]

Most tests are passing locally, all except 3. That's a non-issue; I'm testing that errors get logged, and can't restart docker in the middle of the tests. If I could, those tests would pass.