perf: reduce sketch hashing overhead

[tisonkun]

Summary

• improve hash-backed sketch update performance for integer and raw-byte inputs
• preserve existing MurmurHash3 and XXHash64 outputs and streaming behavior
• add end-to-end update benchmarks for every sketch using these hashers

Design Notes

The optimization keeps the public API and hash algorithms unchanged. It focuses on the existing hot path from a sketch update through Hash and the internal hasher, making that path easier for the compiler to specialize while keeping full-word processing separate from partial tails.

The benchmarks exercise public sketch APIs rather than isolated hash helpers. Integer inputs benefit most from exposing the complete fixed-size hashing path to the optimizer. Raw-byte inputs additionally benefit from exposing the single contiguous write and using fixed-width reads for complete words. Count-Min performs four seeded MurmurHash3 passes per update in this benchmark, while Bloom performs two XXHash64 passes, so reductions in per-hash overhead accumulate in those sketches.

Benchmarks

Local Divan benchmark medians for 10,000 updates:

Sketch	Input	Before	After	Change
Bloom	u64	383.2 us	303.5 us	20.8% faster
Bloom	32-byte raw input	513.5 us	398.6 us	22.4% faster
Count-Min	u64	268.5 us	70.53 us	73.7% faster
Count-Min	32-byte raw input	281.9 us	242.0 us	14.2% faster
CPC	u64	128.8 us	89.42 us	30.6% faster
CPC	32-byte raw input	144.4 us	139.8 us	3.2% faster
Frequent Items	u64	338.8 us	129.4 us	61.8% faster
Frequent Items	32-byte raw input	397.0 us	388.1 us	2.2% faster
HLL	u64	148.2 us	96.06 us	35.2% faster
HLL	32-byte raw input	174.4 us	143.9 us	17.5% faster
Theta	u64	120.1 us	49.57 us	58.7% faster
Theta	32-byte raw input	120.0 us	108.9 us	9.3% faster
Tuple	u64	125.3 us	56.78 us	54.7% faster
Tuple	32-byte raw input	124.5 us	116.8 us	6.2% faster

The baseline uses the same benchmark code with datasketches/src/hash restored from main; the optimized run uses this PR. Results are machine-dependent and document local before-and-after evidence rather than a portable performance guarantee.

Validation

• cargo x check
• cargo x test
• cargo x lint
• cargo bench --package benchmarks --bench benchmarks -- bloom::update countmin::update cpc::update frequencies::update hll::update theta::update tuple::update --min-time 1 --sample-count 20