Speed up sharding optimizer build

[AlbedoWang]

Summary

This PR reduces AutoParallel sharding optimizer build time while preserving the solver-visible optimization problem and selected placements.

The main changes are:

• Skip enumeration-time DTensor redistribute_cost work during sharding metadata construction. Those costs are not consumed by the solver directly; AutoParallel recomputes and stores its own edge costs before building the ILP objective.
• Avoid creating PuLP variables for infinite-cost strategy edges. These edges were previously represented as variables and then constrained to zero; now they are omitted before variable creation while constraints treat missing variables as zero.
• Store repeated-subgraph cluster links at node granularity instead of per (node, arg, output, input) option. Identical repeated nodes resolve to their root node using matching option indices, reducing cluster bookkeeping size.
• Add deterministic parallel edge-cost construction controlled by AP_PARALLEL_BUILD, with serial fallback when CUDA has already been initialized.
• Add AP_FAST_BUILD=0 as an A/B switch for disabling the enumeration-time shortcut.
• Update optimizer serialization to preserve the new node-level cluster links and valid decision-variable key set.
• Add user-facing and design docs for fast optimizer build.
• Update tests to cover public end-to-end behavior: fast/reference placement equivalence, JSON export behavior, serialization round trips, and dtype-cast placement behavior.