feat: SDK memory optimization — sketch pooling, series ID fast-path, label heap free

[zzylol]

Summary

• sync.Pool per sketch aggregator: series structs and their sketch backing arrays (KLL compactors, HLL registers, CMS count tables, etc.) are recycled via Reset()/Clear() instead of reallocated each export cycle. Eliminates up to ~197 KB of heap churn per series per delta export for CountMinSketch.
• SeriesIDSink fast-path: after the stateful OTLP Dictionary assigns a series ID on first export, subsequent exports short-circuit before any fingerprint string building or map lookups.
• AttrsClearer: after first registration, the Dictionary zeroes series.attrs via a back-pointer, permanently freeing the []KeyValue heap array (~224 bytes per 5-label series).
• Dictionary memory management: sync.Pool for *strings.Builder and *seriesEntry; generation-based eviction (staleGenerations=5) for stale series; idle-series eviction in cumulative aggregators (maxIdleCycles=3) to handle dynamic label churn.

Sketch aggregators added

Type	Backing array freed per Reset()
HLL	16 KB (2¹⁴ registers)
KLL (k=200)	~25–50 KB (compactor arrays)
DDSketch	~10–30 KB (bucket map)
CountSketch (5×2048)	~82 KB
CountMinSketch (4×2048)	~197 KB

Test plan

• go test ./sdk/metric/... passes
• go test ./sdk/metric/internal/aggregate/... passes
• Build clean: go build ./...
• Integration test with collector end-to-end

🤖 Generated with Claude Code