core: optimize AllocsFit to minimize calculation

[schmichael]

AllocsFit is called from the plan applier and binpacking iterator to
ensure planned/proposed allocs fit on a node. While it always compares
mem/cpu resources, it does not always need to check devices and
networks.

• Plan Applier

  • Before this change the plan applier would always check devices and
    networks for collisions.

  • After this change the plan applier will only checks devices and
    networks for collisions if the plan's job contains devices and
    networks. There's no possibility of a collision if there are no new
    device or network reservations.

• Binpack Iterator

  • Before this change the binpack iterator would never check ports
    or devices because those feasibility checks are handled elsewhere.
    Bandwidth was checked but hardcoded to pass since Nomad 0.12.

  • The behavior has not changed, but the code is much cleaner. Before a
    network index would be passed to AllocsFit, but only its bandwidth
    would be checked. Since bandwidth is deprecated, this check was a
    noop. All of that code is removed so it no longer appears that
    AllocsFit is checking for network collisions when it does not.

TODO

• Tests!
• Benchmark
• NetworkIndex usage could probably still be improved.

[dadgar]

@schmichael May be nice to add a benchmark to this so before and after can be compared

[hashicorp-cla]

All committers have signed the CLA.