polars: tag Hydra baseline subset keys by column (#74)

[zzylol]

Closes #74. This is PR 1 of the Hydra measurement plan in #193 (docs/rqe_optimizer_hydra.md §5.2).

Problem. asap_sketchlib 0.3.0's Hydra keys each label subset by column (label0:abc;label1:x), so equal values in different label columns never alias in the sketch. The polars exact baselines still built 0.2.2's untagged key (abc;x). On label columns that share a value domain, such as configs/datagen/hydra_columns.yaml's shared alphabet, they merged a key1 group with the key2 group of the same value. As a result the exact rows scored nonzero.

Change

• polars_shared::subset_key builds the 0.3.0 subkey over hydra_shared's label{i} schema, escaping \, : and ; as the library does.
• prefix_key is the lookup for a probe of the leading columns, and every polars lookup uses it: frequency, cardinality, univmon and hydra-kll.
• An empty label keeps its column, as 0.3.0 does. The old key dropped it.
• The two datagen yaml headers no longer describe the aliasing as current.

Tests

• subset_keys_tag_each_value_with_its_column.
• a_value_shared_by_two_columns_is_counted_once: records a;a and a;b. Group a in column 0 holds 2 records; the old key counted 3.
• cargo test -p sketch-bench passes (98), as do fmt and clippy.

Evidence. These are the polars rows on the shared-alphabet data: hydra_columns.yaml, and its _f64 version for KLL. Grids: cms 3×128 with 3×512 cells, hll 3×128, kll 3×64 with k=200. --operations query --metrics accuracy.

Row	main (f9c4b68)	this branch
hydra-cms polars	are_all 0.596, aae_all 3.74	0, 0
hydra-hll polars	are_all 0.249, aae_all 56.8	0, 0
hydra-kll polars	mean_rank_err 0.00093, max_rank_err 0.036	0, 0

🤖 Generated with Claude Code

https://claude.ai/code/session_01FhcWJExZcmVqS6r6rEtjis