perf(strscan): make the compile cache thread-safe (sync.Map)

[tannevaled]

What

The compiled-regexp cache that lets a repeated pattern source compile only once — instead of recompiling via onigmo.Compile on every Scan/Match/Skip/Check (and *Until) — was backed by a plain, unsynchronized map[string]*onigmo.Regexp. Caching by the source string is correctness-safe (the flags a caller wants are embedded inline in the source, e.g. (?imx)..., so the compiled regexp for a given source is deterministic), but the bare map raced under concurrent Scanners: go test -race reliably reports a data race when many goroutines scan at once, and a concurrent map write can crash the process.

The fix

Replace the map with a process-wide sync.Map, consulted at the single onigmo.Compile site:

• Load on the read-mostly hot path is lock-free.
• LoadOrStore collapses concurrent compile-misses of the same source to one shared *Regexp (losers drop their duplicate), so every Scanner reuses the same compiled value.
• Compile errors are not cached and are still surfaced as "no match", keeping error behavior byte-for-byte identical to compiling every call.

Results are unchanged — caching is purely internal. The shared *Regexp's Match is itself concurrency-safe, so Scanners can share one compiled pattern.

Tests added

• strscan_bench_test.go — a lexer-shaped scan loop, cached vs recompile-every-call, to quantify the win.
• TestCompileCacheConcurrent — many Scanners, many goroutines: the permanent -race regression guard (fails on the old map, passes on sync.Map).
• TestCompileCacheMalformedNotCached — a malformed pattern keeps missing identically, and a later valid pattern still compiles.

Benchmark (Apple M4 Max, go1.26.4, CGO_ENABLED=0)

Full tokenizing pass over a representative input with a small fixed pattern set:

benchmark	ns/op	B/op	allocs/op
ScanLoopCached (this PR)	~611,771	1,095,379	28,120
ScanLoopNoCache (recompile every Scan)	~27,846,799	46,283,964	669,327

The compile cache is ~45x faster and uses ~24x fewer allocations than recompiling on every Scan, with zero behavior change.

Conformance

• 100% coverage maintained (compile 100%, total 100%).
• gofmt / go vet / go test -race clean; CGO_ENABLED=0.
• Builds on all 6 arches (amd64, arm64, riscv64, loong64, ppc64le, s390x).
• The MRI-4.0.5 differential oracle still passes against real ruby 4.0.5.

🤖 Generated with Claude Code