feat: add top-level CompareAndSwap and CompareAndDelete (#14)

[millerjp]

Summary

Adds top-level generic CompareAndSwap and CompareAndDelete functions, wrapping the corresponding sync.Map methods with a tighter V comparable constraint. SyncMap[K, V any] stays unchanged — non-comparable V types remain supported, and CAS operations reject them at compile time rather than panicking at runtime.

Closes #14.

Signatures

func CompareAndSwap[K, V comparable](m *SyncMap[K, V], key K, old, new V) (swapped bool)
func CompareAndDelete[K, V comparable](m *SyncMap[K, V], key K, old V) (deleted bool)

Verification

$ go doc github.com/axonops/syncmap CompareAndSwap | head -4
func CompareAndSwap[K, V comparable](m *SyncMap[K, V], key K, old, new V) (swapped bool)
    CompareAndSwap swaps the old and new values for key if the value currently
    stored in m is equal to old. The swapped result reports whether the swap was
    performed.

$ make check
… (all 14 gates green, coverage 100%, no vulnerabilities)

$ go test -race -count=10 -run 'TestCompareAnd' ./...
ok github.com/axonops/syncmap 1.290s

$ go test -bench='CompareAnd' -benchmem -run=^$ -benchtime=100000x ./...
BenchmarkCompareAndSwap-8            115.9 ns/op   56 B/op   1 allocs/op
BenchmarkCompareAndSwapMismatch-8     42.6 ns/op    7 B/op   0 allocs/op
BenchmarkCompareAndDelete-8           82.7 ns/op   48 B/op   1 allocs/op
BenchmarkCompareAndSwapParallel-8     46.7 ns/op   13 B/op   1 allocs/op

The 1 alloc/op on the success paths is the V→any boxing cost at the sync.Map boundary — the same cost raw sync.Map users pay. Target of "zero additional allocations beyond sync.Map itself" met.

Agent gates

• Plan — one-commit scope, placement in syncmap.go between Delete and Range.
• code-reviewer (pre) — approved; confirmed [K, V comparable] sugar semantics, named-return consistency, single-file rule.
• test-writer (pre) — produced the initial test block; id starts at 1 so the stored-0 seed can't be mistaken for a winner.
• security-reviewer (post) — approved with two IMPORTANTs (added godoc note on interface-V runtime panic; added comment on id-range invariant in the contention test).
• performance-reviewer (post) — initially flagged missing benchmarks as BLOCKING; minimal CAS benchmarks added (syncmap_bench_test.go). Full public-method benchmark coverage remains with test: add benchmarks, runnable examples, and bench-regression infrastructure #15 as originally scoped.
• go-quality — 8/8 PASS.
• commit-message-reviewer — PASS.

Deferrals (tracked elsewhere)

• CHANGELOG "Added" entry — docs: add README, CHANGELOG, NOTICE, SECURITY, and LLM docs bundle #17 (CHANGELOG owner).
• BDD scenarios for CAS — test: add BDD (godog) suite and fuzz targets #16.
• Full benchmark matrix (every public method, raw-sync.Map overhead comparison, bench.txt baseline, benchstat-regression-guard) — test: add benchmarks, runnable examples, and bench-regression infrastructure #15.

Test plan

• CI green
• Coverage remains 100%
• go doc CompareAndSwap renders the interface-V caveat