fix: prefer accelerators in heterogeneous placement

[abendrothj]

Motivation

Backend compatibility alone can leave a CPU-only singleton tied with an accelerator singleton. The existing free-memory tiebreaker can then assign the full model to the CPU while CUDA sits idle.

Changes

• score valid candidate cycles by accelerator-backed nodes
• rank Metal/CUDA capability ahead of download locality and free-memory tiebreakers
• preserve existing smallest-cycle, required-node, backend, and RDMA filtering
• add a regression test where a lower-memory CUDA node beats a higher-memory CPU node

Testing

• uv run pytest src/exo/master/tests/test_placement.py -k "accelerator or backend"
• uv run basedpyright src/exo/master/placement.py src/exo/master/tests/test_placement.py
• uv run ruff check src/exo/master/placement.py src/exo/master/tests/test_placement.py

Fixes #2180

---

Refiled from #2213, which was closed unintentionally.