Estimate flex_local_map alternative costs automatically

[AlbedoWang]

Summary

This PR stacks on #518 and adds automatic analytical cost estimation for
flex_local_map alternatives that do not provide a cost_hint.

• Trace each alternative body with its declared placement-local tensor shapes.
• Export and normalize its joint graph through the same AutoParallel path.
• Reuse the existing per-op FLOPs/roofline and unified collective latency models.
• Add dynamic-shape support needed by the DeepSeek V3 MoE EP/no-EP alternatives.

An explicit finite, non-negative cost_hint continues to override automatic
estimation for that alternative. auto_cost=False preserves the zero-default
behavior from #518.

Stack

• Base PR: Add flex_local_map: multiple sharding alternatives for the AutoParallel solver #518
• Base commit: b1f329f
• Automatic estimation: 6a54a68
• DSv3/MoE extension: d5d9a69

Validation

Executed in the repository's autop-ci-py312 conda environment:

PYTHONPATH=$PWD conda run -n autop-ci-py312 \
  python -m pytest -q tests/test_optimize_placement.py

PYTHONPATH=$PWD conda run -n autop-ci-py312 \
  python -m pytest -q tests/test_placement_options_utils.py

Results:

• tests/test_optimize_placement.py: 49 passed
• tests/test_placement_options_utils.py: 30 passed

These are analytical-model checks rather than hardware performance measurements.
No latency, throughput, memory, speedup, or measured model-accuracy claim is made.

Authored with Claude.