hllplus: allocate sparse delta slice and buffer lazily

[antoxas1986]

Problem

The sparse HLL++ state pre-allocates its delta slice and buffer map for the worst case when a sketch is constructed. In streaming/aggregation pipelines that key HLL sketches by a high-cardinality grouping key, many low-cardinality sketches are held live simultaneously, and this worst-case preallocation dominates heap — most of those bytes are never touched because each sketch only ever sees a handful of distinct values.

Change

Two allocations in newSparseState switch from worst-case sizing to lazy/minimal sizing:

• the delta slice starts at a small capacity instead of maxDataLen,
• the buffer map starts empty instead of being pre-sized to maxBufferLen.

The delta slice is append-grown and the buffer is a map, so both grow correctly on demand. maxDataLen/maxBufferLen are retained as flush thresholds — only their use as allocation sizes is dropped. Serialization output and cardinality estimates are unchanged; only the initial per-sketch heap footprint drops.

Benchmark

Low-cardinality sparse sketch (construct + add 10 distinct values), -benchmem -count=5:

	B/op	allocs/op
before (worst-case sizing)	22298	10
after (lazy sizing)	520	9

Tests

• Existing suite passes (go test ./...).
• Added TestLazyAllocInvariant: builds independent sketches over identical input across cardinalities {1, 10, 100, 10k, 1M} at precisions 12/17 and 15/20, asserting byte-identical serialization, identical estimates, and a serialize→deserialize round-trip that preserves the estimate.

[dim]

thanks!!