Add split-dim seed search

[AlbedoWang]

Mesh-discovery sweeps can make full strategy enumeration expensive because every mesh dimension multiplies the placement search space. This PR adds a small split-dim seed primitive: solve one one-dimensional sharding problem per mesh dimension, stitch the per-node placements into a full-mesh seed, then let the normal full-mesh optimizer search only a Hamming ball around that seed.

The seeded path keeps the existing optimizer lifecycle. Strategy generation applies the Hamming-ball restriction, then the usual PuLP decision variables, costs, and constraints are built. get_solution() still solves the restricted problem as an ILP, and the new solve_lp_relaxation() solves the same restricted PuLP problem continuously and reports objective/status/fractionality diagnostics, with solution extraction when the relaxation is integral.

For fabric-aware seed generation, each one-dimensional solve can reuse the NCCL topology derived from the corresponding full-mesh dimension. NCCLTopoConfig.mesh_dim_topo_override is limited to 1D dim0 so this override cannot silently change the general topology derivation path.

Review order

1. autoparallel/mesh_search.py — build_split_dim_seed(...), one-dimensional solve cache keys, fabric-aware per-dimension cost model handling, and seed construction for FX nodes plus input/output companion nodes.
2. autoparallel/shardings/propagation_rules.py and autoparallel/shardings/placement_options.py — active seed state, Hamming-ball filtering, and cache-key isolation so seeded and unseeded placement options do not reuse each other.
3. autoparallel/optimize_sharding.py — strategy_seed / strategy_radius wiring during metadata construction and the LP-relaxation solve path.
4. autoparallel/cost_models/nccl_cost_model.py — one-dimensional mesh topology override for fabric-aware seed solves.
5. Docs and tests — split-dim seed docs plus tests/test_split_dim_seed.py covering seed construction and restricted ILP/LP solve behavior.

Two compatibility fixes are included because they are exercised by the same CUDA CI matrix: local_map HOP target names are accepted by family instead of the older exact call_local_map* spelling, and the fake CUDA device properties used by tests include L2_cache_size for newer Inductor autotuning code.

Test plan

• Lint checks matching CI: isort, black, mypy, flake8
• CUDA 13 nightly pytest tests — 517 passed, 1 xfailed, 4 xpassed
• CUDA examples: example_autoparallel.py, example_llama3.py, example_local_map.py
• 4-GPU CUDA checks: tests/test_dcp_roundtrip.py -v and torchrun --standalone --nproc-per-node 4 examples/example_ds3_local_map.py

Authored with Codex.