[yummybomb]

Stacked on #511. Review that first; this PR only adds the last commit.

Summary

In mixed-size (heterogeneous) mode, each vGPU occupies a fixed run of its GPU's placement slots, and each profile may only start at the driver's heterogeneous placement IDs. The driver puts a new vGPU at the lowest free placement for its type. On an L40S that scatters 2Q across the runs 4Q needs:

• Of the 16 2Q placements, 8 (0, 10, 12, 22, 24, 34, 36, 46) each overlap one 4Q placement, and 8 (4, 6, 16, 18, 28, 30, 40, 42) overlap none.
• Lowest-free puts every other 2Q on an overlapping placement, so eight 2Q leave room for four 4Q instead of eight.

Change

• After configure writes and verifies current_vgpu_type, and before any VM attaches, packPlacement moves the vGPU by writing placement_id.
• Choice: the free placement that blocks the fewest free placements of larger profiles. Sizes are compared smallest first. A tie keeps the driver's placement, so a write only happens for a strictly better one.
• Placement data:
  • Free placements are computed from the current_vgpu_type and placement_id of the GPU's other configured VFs.
  • The table and mode come from nvidia-smi vgpu -s -v -i <GPU> and nvidia-smi -q -i <GPU>, read the same way as the infra role's heterogeneous-mode check.
  • Both are cached per GPU for 10 minutes. A failed load is cached too, so a host without nvidia-smi doesn't run it on every claim.
• Best effort: packing never fails a claim. Equal-size mode, nvidia-smi failing, an unreadable sibling or one with an unknown type, a refused placement_id write, or a read-back mismatch each log and keep the driver's placement.
• This only runs when the type is actually written. In Place vGPU profiles the driver hides from creatable lists #511 that happens inside claimVGPU under the allocation lock, so siblings can't change underneath it.

Validation

• Unit tests (the parser reads the real nvidia-smi vgpu -s -v output from a dev L40S, in testdata):
  • Choice: simulating the driver's lowest-free rule plus packing, eight 2Q land on 4, 6, 16, 18, 28, 30, 40, 42 and leave all eight 4Q placements free. Without packing only four stay free. Each further 2Q then costs exactly one 4Q placement. A tie keeps the driver's placement.
  • configure: writes placement_id and loads the table once per GPU. It keeps the driver's placement in equal-size mode, when nvidia-smi fails, when a sibling's type is unknown, and when the driver refuses the placement write.
• Mutation checks: removing each guard makes at least one test fail. The guards are the call in configure, keep-on-tie, the mode check, the unknown-sibling check, the cache, counting only free larger placements, excluding the VF itself, and only choosing free targets.
• Test runs: go test ./lib/devices/ and go test -race ./lib/instances/ -run 'VGPU|VendorVFIO' pass. lib/devices and lib/instances build for darwin.
• On a dev L40S (driver 580.178.05), I ran this branch's devices.ConfigureVGPU and DestroyVGPU against an empty card, with no VMs attached:
  • Eight 2Q writes: the driver placed each at 0, and packing moved them to 16, 18, 28, 30, 4, 6, 40, 42.
  • Eight 4Q writes then landed at 0, 8, 12, 20, 24, 32, 36 and 44, so the card held 16 vGPUs (all 48 GB).
  • A 17th write was refused with EIO, which is the errno a full card returns for current_vgpu_type.
  • All VFs were reset afterwards.
• Not tested: booting a VM on a VF whose placement was moved. The write happens before attach, and the driver accepted every move, but a boot on dev should confirm it before merging.

Notes

• No extra 2Q room: packing changes where vGPUs go, not how many 2Q a card holds. Mixed mode still allows at most 16 2Q per card.
• Placement choice beyond 4Q: the score also counts larger profiles nothing uses today, such as 8Q and 16Q. They only break ties among placements that are equal for 3Q and 4Q, which is why the dev run moved the first 2Q to 16 rather than 4.