perf(set): single-probe ttak_set_add (and fix inverted presence check)

[gg582]

Summary

ttak_set_add did a full get probe and then a full put probe per call — two hash computations and two table walks for every insert. While reading the code I also found the presence check was inverted: keys were only re-stored when already present, so new keys were never added. This is a live bug — tutorials/07-set-container prints contains beta? no immediately after adding "beta" on current main.

Changes

• src/ht/table.c: extract the insert logic into a shared table_insert() core with a replace flag; ttak_table_put() keeps its exact old behavior, and a new ttak_table_put_if_absent() performs a single probe and reports whether it inserted.
• internal/ttak/table_internal.h (new): declaration of ttak_table_put_if_absent, following the existing internal-header convention (e.g. mem_internal.h).
• src/container/set.c: ttak_set_add now calls ttak_table_put_if_absent — one probe per add, duplicates stay no-ops (existing value never replaced, no destructor invoked).
• bench/bench_set_add_dup.c (new): the A/B benchmark used for the numbers below.

Benchmark (A/B against main built in a clean worktree)

Workload: prefill a set with 100k keys via ttak_table_put (identical table state for both builds, since old ttak_set_add cannot insert), then 1M ttak_set_add attempts — 90% duplicates, 10% rotating fresh keys. gcc -O3 -march=native, best of 5 repeats, interleaved runs:

build	ns/add (run 1)	run 2	run 3	run 4
main (old)	29.81	30.31	29.81	29.78
this PR	22.10	27.58*	21.96	22.05

* single outlier run; all other new-build runs cluster at 22.0–22.1. ~26% faster, stable across runs, well above the 5% bar. The win comes from halving probes on duplicate-dominant adds.

Verification

• make test (gcc): all tests pass.
• Tutorial 07 now prints the documented output (contains beta? yes, contains alpha? no after removal).

Notes

• ttak_table_put semantics are unchanged (replace + val_free on existing values), so ttak_map and all other table users are unaffected.
• Remaining known candidates (pool allocator division, pocket remote-lock) were left untouched — this was the clearest win.