speed up formatting and pattern compilation

[lestrrat]

• collapse zero-padded %H/%I into stdlib layout so the whole pattern combines into one time.AppendFormat call (cached path -23%)
• index the specification set by a [256]Appender array instead of a map, dropping per-lookup hashing
• append verbatim text inline on the uncached compile path, cutting Format() from 15 allocs to 1
• cache compiled patterns in package-level Format() (bounded sync.Map, default-spec only), making repeated calls ~4x faster

Benchmarks (AMD Ryzen 9 7900X3D, go1.26.1): package-level Format() 820→206 ns/op (15→2 allocs); cached FormatBuffer 167→137 ns/op (0 allocs). All variants now beat ncruces/fastly/jehiah/tebeka.