Issue sweep 25.08: flashinfer in Setup, community results, shm ghosts, gotcha 41

- Setup installs flashinfer-python + flashinfer-cubin==0.6.13 and verify.sh
  now tests has_flashinfer() instead of a bare import, which passes while the
  DFlash2 selector silently falls back to torch.topk at half speed (#35).
  Launchers export FLASHINFER_DISABLE_VERSION_CHECK=1 for the cubin pair.
- README: results-from-other-hardware section — 4090 harness row (#32),
  170HX sm80 and 5090 as method-flagged notes (#27, #35), dual-GPU links.
- Launchers unlink stale /dev/shm/vllm_offload_*.mmap regions no live process
  maps; a dead engine + restart policy otherwise boot-loops on OSError: Bad
  address, 70 ghosts deep (#33). VLLM_OFFLOAD_KEEP_SHM=1 opts out.
- Gotcha 41: sm80 VMM churn wedges the card in Marlin repack; ahnguyen17
  root-caused it in #27 with a bit-exact CPU-repack workaround.