fix: chunk series in parallel forecast

[jmoralez]

Submits chunks of series to each executor to reduce the communication overhead. The behavior is to create 100 tasks for each executor, so that we can better balance the work whilst keeping communication low and update the progress bar relatively frequently.

The relevant changes in behavior are:

• The progress bar will now show the number of tasks (100 * n_jobs) instead of the number of series. It also won't show the rate and the expected remaining time, since it varies too much and can be confusing.
• The overhead of the parallel processing is very low again (as in 1.7.5), so simple models like the SeasonalNaive run very fast. On my tests with SeasonalNaive this is slightly faster than 1.7.5.
  • 500k: 1.7.5: 16s, this PR 13s
  • 1M: 1.7.5: 30s, this PR 25s

Fixes #898

[review-notebook-app]

Check out this pull request on 

See visual diffs & provide feedback on Jupyter Notebooks.

---

Powered by ReviewNB