replace basestring with six.string_type for python 2/3 compatibility

PY3 support

Thanks to @berriz for raising the issue. Fixes chapmanb/bcbio-nextgen#1427.

PBSPro cannot schedule uninterruptable array jobs, so instead of running the
engines as an array, we create a single engine job submission script and submit
it repeatedly for each job.

This commit also removes some dead code for dill support.

Thanks to Endre Sebestyen (@razZ0r), Francesco Ferrari and Tiziana Castrignanò
for raising the issue, providing an account with access to PBSPro on the Cineca
cluster and testing that the fix works.

We were running into: https://github.com/ipython/ipython/issues/9281

Do not pin ipython so hard

Update README.rst

SGE's parallel environment specification is space-delimited and UGE is comma delimited.
This handles fix handles both. Other than that it seems to work fine.

Thanks to Andrew Oler for spotting this issue, figuring out what wrong, suggesting
the fix and testing the fix.

https://github.com/ipython/ipython/issues/9159

If we wanted 8 GB per engine, 1 engine on the machine with
4 cores for each engine, and the resources are per slot (core)
we should be asking for 8 GB/4 cores = 2 GB/core per engine.

If we are jamming 10 engines on the machine and we want 8 GB
per engine and 4 cores for each engine and resources are per
slot (core) then we need a total of 80 GB on the machine,
and 40 cores so we should be asking for 2 GB/core.

Before mem/self.cores for the 10 engine instance was asking
for 8gb/4 cores = 2 gb/core. But then we were multiplying it
by the number of engines, so we were asking for 20 gb/core.

More frequent heart monitor pings coupled with a larger number of misses
enables an hour of interruptable time but quicker responsiveness on
startup. Thanks to @mwojcikowski fixes #32

Passing --signal=KILL leaves some jobs of the job array hanging
around, so going back to just cancelling the jobs.

Closes #31. Thanks to @mwojcikowski for finding the cause and
a workaround.