[CI] Check B200 NUMA mapping against sysfs numa_node

[mmangkad]

Summary

TestB200NumaTopology hardcoded a one-NUMA-node layout. Cirrascale B200 hosts give each PCIe root complex its own node, so the 4 visible GPUs span two:

AssertionError: 2 != 1 : Expected all visible 4-GPU B200 devices on one NUMA node, got {3, 5}

Failing job

	before	after
asserts	all 4 GPUs on one node	NVML affinity contains the GPU's sysfs numa_node
oracle	hardcoded	/sys/bus/pci/devices/<bdf>/numa_node
BDF from	n/a	torch.cuda, not NVML
BIOS-independent	no	yes

NVML is avoided for the BDF because resolving that handle needs _get_nvml_device_index, whose device_id fallback is one of the faults this test detects.

Host layout (cirrascale1 and cirrascale3 identical, 6 NUMA nodes):

GPUs	NUMA node
0,1	0
2,3	2
4,5	3
6,7	5

Still fails on: device index returned instead of node, bitmask decode offset.

Verification

Test run in each B200 CI container, 4 GPUs compared each, no skips:

container	vendor	GPU -> NUMA node	result
b200-cirrascale1-0123	Cirrascale	0,0,2,2	Ran 1 test ... OK
b200-cirrascale1-4567	Cirrascale	3,3,5,5	Ran 1 test ... OK
b200-cirrascale3-0123	Cirrascale	0,0,2,2	Ran 1 test ... OK
b200-cirrascale3-4567	Cirrascale	3,3,5,5	Ran 1 test ... OK
b200-di01-0123	DeepInfra	0,0,2,2	Ran 1 test ... OK
b200-di01-4567	DeepInfra	3,3,5,5	Ran 1 test ... OK
b200-fin03-4-0123	Verda	0,0,0,0	Ran 1 test ... OK
b200-fin03-4-4567	Verda	1,1,1,1	Ran 1 test ... OK

The Verda hosts have 2 NUMA nodes and are the layout the old assert was written for; they stay green.