ENH implement local fit for trend curve

[BorisMuzellec]

Reference Issue or PRs

Fixes #231

What does your PR implement? Be specific.

Implement a "local" option for the trend curve using scikit-misc's loess implementation.

A few issues to work on:

• scikit-misc loess objects are unpicklable. As a result, using a "local" curve fit makes the DeseqDataSet object unpicklable.

• On some platforms (Ubuntu in particular), loess will cause a segmentation fault, which is not possible to catch using python try ... except .... As a workaround, loess is run in a separate process, to avoid the main process crashing in case of a segfault. This is a bit cumbersome, because the child process cannot communicate the fitted loess object to the main, as it is unpicklable. The current solution is to

  1. Launch loess in a child process
  2. Look at the child process' error code
  3. If the child process terminated correctly, run loess in the main process