feat(planner): infer label groups from series dataset

[milindsrivastava1997]

Summary

Implements #693, the offline dataset-backed path for inferring label-group cardinality N_g.

• Adds a CSV series-inventory loader: one row per unique metric series, with a fixed metric column.
• Derives per-metric schemas and per-profile distinct grouping counts after exact label filters.
• Requires --dataset for the greedy optimizer CLI and fails loudly on missing metrics, labels, filters, empty matches, schema drift, duplicates, and unused columns.
• Threads integer counts into candidates and applies them only to non-subpopulation-aware ingest/query costs.
• Keeps dataset parsing separate from optimizer core so Optimizer: auto-infer N_g from Prometheus #525 can later supply an equivalent Prometheus-backed profile provider.

Verification

• cargo fmt --all -- --check
• cargo clippy -p asap_planner --all-targets -- -D warnings
• cargo test --workspace (all tests passed)

Sketch-bench PR 123 adds explicit, bounded timestamped external trace adapters and metadata. It is complementary infrastructure; this PR consumes a static series inventory for planner cardinality and does not couple the planner to those trace formats.

Part of #525.