Produce containers topology from endpoints via processes

[tomwilkie]

With this change, the process topology is used to produce the containers topology - fixes #228.

Changes:

• render.Map -> render.LeafMap (Leaf to indicate that its a tree of renderers, and these ones have to be at the leaf of that tree)
• a new render.Map, which allows for mapping from NMD -> NMDs. Tests for this.
• add and accumulate some NodeMetadata in the RenderableNode struct, to enable the above remapping
• Some WIP mapping function for identity mappers and mappers to up-convert (ie from endpoints to processes)

The meat of the review is the new render.Map type, the mapping functions and render/topologies.go (naming suggestions?).

Left TODO:

• remove some annotations on nodes from the probe
• preserve pseudo node behaviour with the new mappers
• filter out empty nodes from the process topology
• moar testing

Note I won't be going the container image topology in this PR, so "the old way" of producing renderablenodes will be around for a little longer.

[peterbourgon]

From discussion:

• Every Leaf mapping MUST append the origin node ID to the list of OriginNodes in the destination RenderableNode. This is done in LeafMap.Render.
• Identity LeafMapFuncs (mapping into the same domain, e.g. process to process) SHALL copy their NodeMetadata into the destination RenderableNode. This is done in NewRenderableNode, which is always invoked by the identity LeafMapFuncs.
• Non-identity LeafMapFuncs (mapping from one domain into another, e.g. endpoint to process) MUST NOT copy their NodeMetadata into the destination RenderableNode. Absent a sophisticated, per-key metadata merge strategy, that copy would be both lossy and destructive. Further, that metadata isn't necessary for building the topology, and can be retrieved later by following the OriginNode foreign keys. This is all currently true, as non-identity LeafMapFuncs always invoke newDerivedNode, which doesn't touch NodeMetadata.

[tomwilkie]

@peterbourgon I think this is ready to go now. You were right about not-propagating-NMDs for derived nodes.

[peterbourgon]

OK let me do another parse...

[peterbourgon]

Okay, I follow everything. After those edits, LGTM.

[tomwilkie]

Address feedback; thanks!

there are 2 problems with the PR as it stands:
(1) containers by image is broken; I have another PR for that. Will "fix it up" in this branch
(2) we get this weird periodic "shuffling" in the UI - I think its something to do with the process topology getting out of sync with the endpoint topology