Expose and consume finer-grained pcie root locality data in the cpu allocator

[ffromani]

Raised in #163 (comment)

We need to:

1. consume the pcieRoot locality data in the cpu allocation algos - which requires us to go and effectively fork the cpumanager code we import and pseudo-vendor.

2. actually having fine grained pcie Root data to feed the allocator into - this is tied to the previous overlay idea which I proposed and then ditched in expose the pcieroot device attribute finding the roots from devices #156 .

Feeding the allocator with finer-grained data is actually pretty-easy. We need to revamp the overlay concept originally added in #156. #163 merged and gives us a simple and straightforward way to feed the data, fully opt-in, using the sysfs API.

The main challenge is how to consume sub-NUMA locality data. This is also tied to the LLC affinity data. Currently we mostly borrow the cpumanager allocation code, its limitations and design decision. A prerequisite for this work is to fork that codebase.

Once we have a custom allocator which can take into account sub-NUMA locality, we can fix the pcie root locality problem and then iterate over the LLC locality data, providing better solutions.