feat: support manual pipeline layer allocation

[abendrothj]

Motivation

Memory-proportional automatic allocation cannot express deliberate asymmetric load, such as assigning only a few layers to a laptop while a dedicated larger machine carries the rest.

Changes

• add optional node_layers to the placement API and PlaceInstance command
• require the supplied nodes to exactly match a connected pipeline cycle
• validate positive counts, exact model-layer totals, and per-node memory capacity
• reject manual allocation for tensor and CFG-parallel modes where layer counts do not map one-to-one to physical nodes
• generate pipeline shard boundaries from the requested per-node counts
• preserve automatic proportional placement when node_layers is omitted

Example request field:

{
  "node_layers": {
    "small-node-id": 2,
    "large-node-id": 38
  }
}

Testing

• uv run pytest src/exo/master/tests/test_placement.py src/exo/master/tests/test_placement_utils.py src/exo/api/tests -m "not slow" (94 passed)
• repository-wide uv run basedpyright
• repository-wide uv run ruff check

Fixes #1720

---

Refiled from #2214, which was closed unintentionally.