gen: optimize unconstrained byte array serialization

[gabrix73]

Summary

• emit one bounds check and copy when parsing unconstrained u8 and char arrays
• append byte slices directly when encoding and omit empty per-byte validation loops
• retain element-by-element handling for constrained byte types
• regenerate affected fixtures and add benchmarks plus an input-ownership regression test

The parser still copies decoded bytes into owned storage, so mutating the input buffer after parsing does not change the decoded value.

Benchmark

64 KiB leftover byte array on linux/amd64, Intel i5-6300U, median of three 500 ms runs:

Operation	Before	After	Allocations
Parse	208016 ns/op	23563 ns/op	23 to 2
Marshal	171976 ns/op	23572 ns/op	22 to 2

Command:

go test ./gen/tests/rem -run '^$' -bench 'Benchmark(Parse|Marshal)ByteTail$' -benchmem -benchtime=500ms -count=3

Verification

• go test ./...
• go vet ./...
• git diff --check