[analytics-engine] Add ShardBucketOversamplingIT for shard-bucket oversampling

End-to-end coverage for the shard-bucket-oversampling rule against a 2-shard
parquet-backed index.

  - Setting wiring + validation: default factor / factor=0 / factor=1.0 / factor=3.0 /
    invalid-factor (rejected at create-time) / dynamic update.
  - Pass-through path: `stats sum(value) as total by category | sort - total | head 100`
    over a 6-doc index — asserts `{a:5, b:7, c:9}`.
  - Decomposed path: `stats avg(value) as avg_v by category | sort - avg_v | head 100` —
    asserts `{a:2.5, b:3.5, c:4.5}` on both default-factor (rule fires) and factor=0
    (rule disabled) paths.
  - Engine-native-merge path: `stats dc(value) as dc_v by category | sort - dc_v | head 100`
    over an index with 30 distinct values per category — asserts `{a:30, b:30, c:30}` on
    both default-factor (SHARD_MERGE + hll_estimate) and factor=0 (PARTIAL→FINAL) paths.
    HLL is exact above its sparse-mode regime; 30 distinct per category sits well inside
    that regime.

Polling helper waits for `SUM(value) BY category` to stabilize before each assertion fires
(parquet rowgroup commit timing is asynchronous to refresh).

Signed-off-by: Sandesh Kumar <sandeshkr419@gmail.com>

Author: sandeshkr419