perf: purge in chunks of 8, not 100 - #5134

Stacked on #5133.

purgeChunkSize bounds how many documents share one purge transaction. It keeps a commit under the store's transaction size limit, which it does at 100, but it also sets the cost of the purge and that part was never measured.

Each purged document opens several iterators inside the chunk's transaction, and badger rebuilds and re-sorts the transaction's whole pending-write set on every one, so a chunk costs roughly the square of its size and total purge time is close to linear in the constant. It is the same cost mergeChunkSize = 8 was introduced to remove from the merge path.

BenchmarkPurgeByDocIDsChunkSize, added here, measures 2,739 docs/s at 8 against 713 at 100 and 365 at 200, on disk with three secondary indexes and history pruning on. A host runs 44 to 203 docs/s on the same code, so the magnitudes do not reconcile and this is a candidate to A/B rather than a proven fix. Sitting at the top of the stack means the treated build differs from what the test mesh already runs by this constant alone. If drain on a treated host stays under 100 docs/s the cost is in the per-document DAG walk, not the chunking.