Add version-consistent result rounding to load_balance_peers

[justheuristic]

Problem:
In current version, calling load_balance_peers may result in a nasty bug: two peers with different numpy/scipy versions (or even different builds of the same version, such as mkl/nomkl) can make incompatible decisions on how to round load balancing outputs.

Here's an example from my local laptiop (numpy+atlas)

Here's a colab instance with the same version of numpy+scipy, but different build and python version

When training collaboratively, all AWS peers would split vector parts as:
(4461869, 4461869, 4461868, 4461868, 0, 0, 0, 0)

... while a some desktop peers decided:
(4461869, 4461868, 4461868, 4461869, 0, 0, 0, 0)

As a result, AllReduce failed with INTERNAL_ERROR due to incompatible part sizes.

Solution:

• rounded load balancing results up to 10^-9
• hard-code a solver method in case scipy will ever change the default solver (they did so in the past)
• added a callback for leader to solve linprog for the group