perf(dv): avoid dense expansion for sparse masks

[fallintoplace]

Summary

• Clear sparse deletion positions directly in the keep mask.
• Keep the dense word path for dense and run-heavy vectors.
• Skip out-of-range buckets without expanding them.
• Add sparse and dense benchmarks plus a high-position correctness case.

Benchmark

Apple M1 Pro, arm64. The mask covers 1 MiB of rows:

Case	Before	After	Result
10 sparse deletes	95.2 us, 254 KB, 3 allocs	25.0 us, 131 KB, 2 allocs	3.8x faster, 48% fewer bytes
1,000 sparse deletes	108.1 us, 262 KB, 3 allocs	27.5 us, 131 KB, 2 allocs	3.9x faster, 50% fewer bytes
Dense range	115.0 us, 262 KB, 3 allocs	112.4 us, 262 KB, 3 allocs	essentially unchanged
Out-of-range delete	208.3 us, 2.1 MB, 3 allocs	126.6 ns, 120 B, 2 allocs	no dense temporary

Tests

• go test ./...
• go vet .