Add Builder for fast sorted fingerprint insertion

[lukesandberg]

Summary

• Adds a Builder type for constructing filters from fingerprints in sorted (non-decreasing) order. Each insertion is O(1) amortized via sequential append — no run boundary lookups, linear scans, or element shifting.
• grow() and shrink_to_fit() internal paths now use Builder for reconstruction, improving their performance.
• At ~95%+ occupancy when slots wrap around, the builder delegates to Filter::insert_impl for correctness and simplicity
• Out-of-order insertion panics unconditionally, making the sorted contract explicit.

API

let mut builder = qfilter::Builder::new(1000, 0.01).unwrap();
// Also: Builder::new_resizeable(), Builder::with_fingerprint_size()

let fp_size = builder.fingerprint_size();
let mut hashes: Vec<u64> = items.iter()
    .map(|i| qfilter::compute_fingerprint(i, fp_size))
    .collect();
hashes.sort();

for h in hashes {
    builder.insert_fingerprint(false, h).unwrap();
}
let filter = builder.into_filter();

Benchmarks (10k items, criterion change% vs baseline)

Benchmark	Change	Notes
sorted_insert	~92 µs	New — ~7x faster than grow (regular insert to fill)
grow_from_90pct	-11.5%	Internal rebuild now uses Builder
grow_resizeable	-16.7%	Multiple growth cycles, each rebuild faster
shrink	-11.5%	shrink_to_fit rebuild uses Builder
shrink_10pct	-11.8%	Same mechanism
grow	neutral	No growth triggered in this bench

Test plan

• cargo test — 35 unit tests + 6 doc-tests pass
• cargo clippy — clean
• cargo +nightly fuzz run fuzz_sorted_insert -- -max_total_time=120 — 636K iterations, 0 crashes (includes resizable growth paths)
• cargo bench — no regressions, improvements in grow/shrink paths

🤖 Generated with Claude Code

[arthurprs]

Thank you for your contributions. I skimmed through it and it looks good. I'll find more time in the next few days to review and merge.

[lukesandberg]

Thanks! We are excited about getting these changes in and some of your recent perf improvements. They appear to have some pretty dramatic perf wins for our usecase in turbopack.