Modisco

Fix in https://github.com/kundajelab/tfmodisco/releases/tag/v0.5.16.4 reported to not work. Should have done `min(perplexity, matrix.shape[0]-1)` rather than `min(perplexity, matrix.shape[0])`. New fix in light of message on https://github.com/kundajelab/tfmodisco/issues/112

Reported in issue https://github.com/kundajelab/tfmodisco/issues/112

Bug was caused by the addition of a feature added in a later release (subclustering within motifs and visualization of the subclusters using t-sne). Fix is to put in a check to reduce the perplexity relative to the default if the number of seqlets in the motif is less than the default perplexity.

Corresponds to PR https://github.com/kundajelab/tfmodisco/pull/108 by akmorrow13

Corresponds to PR https://github.com/kundajelab/tfmodisco/pull/99

- Removed some tolist() commands that might have been contributing to memory explosion
- More detailed printouts of memory usage
- Made it possible to specify a different number of parallel runs for the main clustering step via the n_cores_mainclustering argument to TfModiscoSeqletsToPatternsFactory

- Added pynnd=True option to use pynn descent for coarse-grained affinity matrix computation (caveat: runs into a weird pickling error on Colab: https://github.com/lmcinnes/pynndescent/issues/133)
- Noticed that storing the agkm embeddings as [(agkm_string_representation, value), ...] seemed to take up a lot of space (possibly because representing the agkms as strings is space-consuming? So now they get converted to [(agkm_idx, value)...] before being stored. This seems to bring down the memory consumption.
- Other minor changes pertaining to reporting some internal hit-scoring-related metrics (exclude_self excludes the self when benchmarking how well the fann_perclass (finegrained-affinity nearest-neighbors) method works for recovering the true class for motif hits, since the fine-grained affinity to the self is always 1; also added benchmarking of how well simply using aggregate similarity works)
- Also did some reorganization of example notebooks that I mainly use to test out stuff - put some of the more experimental notebooks under "examples/simulated_TAL_GATA_deeplearning/other"
- Updating Leiden version to avoid the segfault bug (https://github.com/vtraag/leidenalg/issues/68)

Corresponds to PR https://github.com/kundajelab/tfmodisco/pull/98
- Added circleci continuous integration
- Removed the .travis.yml
- Bumped the version from 0.5.16.0->0.5.16.1
- Added a badge to the github readme
- No tfmodisco code changes