[Backfill corrections] Allocate enough shared memory for parallel prediction generation

[nmdefries]

Description

Main change here is to make sure the docker container has enough shared memory to generate predictions.

Smaller changes:

• Logging
• Make sure receiving directory is cleared after a run
• Check for "SIGBUS" message in log file

Changelog

• Makefile
• main.R

Fixes

Generating predictions in parallel (even with 2 cores) resulted in the error 'memcpy' resulted in a SIGBUS (no shared memory left). Docker containers have 64 MB shared memory by default, which is apparently sufficient for modeling training in parallel but not generating predictions. This change assigns 2 GB of shared memory, which is enough for generating predictions in parallel with up to 3 cores (and maybe more -- didn't test).

[krivard]

👍

nice extension of the log checks; it's always easier on monitoring if the make fails when something has gone wrong