fix how AbstractTrainer and tester apply sigmoid

[NickleDave]

• both should have apply_sigmoid method and check if it is True or False
• and use the specified sigmoid threshold when they do
• instead of behavior depending on size of output of network
  • prefer to throw an error when user-specified arguments cause it, then be able to fail silently because of internal logic

[NickleDave]

both classes now modified to always apply sigmoid when computing multi-label predictions

less of an issue since for VOCDetection dataset we now compute all the predictions + metrics, no need to distinguish which one we're computing

and as the issue says, it is explicit -- behavior depends on loss function, not on size of of output of network

so this issue got resolved another way. closing