Major "waiting model" PR

[rakivo]

It turns out that if the amount of targets is not that big (<100), the waiting model may be faster than the polling one.

For now I'm merging this into master, but on next week rush will automatically decide to use either polling or waiting model, to achieve the maximum speed and efficiency possible.

There's also some minor bug fixes and new features:

• "subrush" keyword. See: Analog of "subninja" #52
• The overwriting output thing. See: Overwrite previously printed commands and their outputs with the new ones #56
• Lazy clean job. See: Generate clean job only if needed #57