Extdev fpgas

[rowleya]

The main change here is to support external devices connected via the FPGAs and the addition of 2D support to vertices. The main highlights:

• Allows a single device to be connected via multiple FPGAs, and different input and output FPGAs
• Adds support for 2D vertices (and other dimensions technically)
• Simplifies support of routing constraints on devices by having a simple method override rather than a constraint
• Removes constraints that all things use anyway and all allocators allocate with anyway
• Removes virtual chips, instead relying on algorithms being aware of how devices work

Depends on SpiNNakerManchester/SpiNNMachine#181

[Christian-B]

Lots of small comments. None are blockers but all should be address at some point (ok for version 2)

[andrewgait]

I'm happy with this; just made one more comment here that might be worth thinking about.

[andrewgait]

I wonder whether calling this variable max_atoms_per_dimension_per_core as well might be a good idea. (I'm not sure whether PyNN has named this variable already and that might mean that might not be a good idea...). I suppose because in the majority of cases it will still be a single int that it's probably not worth changing but I thought I'd put the suggestion out there.

[rowleya]

I did wonder the same thing but didn't for two reasons:

1. It may be hard to find places where this is used as a keyword argument given the number of application vertices around, and how some might be called via super(...).
2. Arguably, max_atoms_per_dimension_per_core returned elsewhere is slightly different, in that it is always a tuple, where this can be a tuple, int or None.