Change Python environment to distribute software

**Fun fact:** In Python 3 you can start Process objects from different Python installations

Example
--------

```python
In [1]: import sys

In [2]: sys.version
Out[2]: '3.6.0 |Anaconda custom (64-bit)| (default, Dec 23 2016, 12:22:00) \n[GCC 4.4.7 20120313 (Red Hat 4.4.7-1)]'

In [3]: import foo

In [4]: foo.f??
Signature: foo.f()
Source:
def f():
print(sys.version)
File: ~/foo.py
Type: function

In [5]: import multiprocessing

In [6]: ctx = multiprocessing.get_context('forkserver')

In [7]: ctx.set_executable('/home/mrocklin/Software/anaconda/envs/35/bin/python')

In [8]: proc = ctx.Process(target=foo.f)

In [9]: proc.start()

3.5.3 | packaged by conda-forge | (default, Feb 9 2017, 14:37:12)
[GCC 4.8.2 20140120 (Red Hat 4.8.2-15)]
```

How to use this
----------------

A long time ago we had a feature where you could distribute a relocatable conda environment to workers and then have those workers restart themselves in that environment. We removed this because it complicated the nannies. This method however may be much simpler. This need has come up twice recently:

1. In Yarn, where we create conda environments and ship them with Yarn
2. In dask-kubernetes, where we ask users to create their own docker containers

Adding some software environment management to dask would help to resolve both of these problems