feat(opt): emit rule comments as options

[3Hren]

A // comment is a rule option, as in FreeBSD, where TOK_COMMENT sits among rule_options and fill_comment writes an O_NOP opcode that keeps its F_NOT and F_OR bits. The parser now hands it to State.OnOption as an OptComment whose Text is the rest of the line up to any #, leading space kept and trailing whitespace removed, instead of storing it in Instruction.InlineComment.

The comment arrives wherever it stands: after a rule body, after check-state, and as the body of a comment-only rule (add 100 // note still gives a count rule with TargetAny source and destination, then the comment option). ParseOptions consumes it instead of stopping at it, and an OptionHook is never asked about //.

Taking the rest of the line, a comment inside { … } leaves the group unclosed: { in or // c } now fails with ErrExpectedOr at the end of the line, where it used to be ErrUnknownOption. not // text is accepted, as upstream does, and makes the rule never match.

VM

program.DropComments removes a comment that decides nothing before a rule is closed: not negated, not in an or-group and not followed by an option joined to it. A commented rule therefore keeps an empty option run and the fast path of a rule without options. A negated comment, or one a hook placed in a group, stays and holds through matchPolicy, so not // fails its rule.

Unchanged divergences

The accepted syntax is otherwise the same as before. FreeBSD requires // to be a whole word, here glued forms such as in// c keep being accepted. The 80-character limit is not enforced and the words are not rejoined with single spaces.

Compatibility

Breaking: Instruction.InlineComment is removed, the comment is an option. v0.2.0 is already incompatible with v0.1.0.

Performance

taskset -c 12-15, prebuilt binaries before and after, benchstat, n=6:

benchmark	before	after
Parser_Next_CommentOnlyRule	124.2 ns	178.8 ns
Parser_Next_CommentOptionLong (about 4 KB of comment)	905 ns	1030 ns
Parser_Next_OptionsAfterTarget	2.33 µs	2.37 µs (p=0.46)
Parser_Next_Grammar/Ruleset/Raw	9.81 ms	10.12 ms (p=0.065)

A comment line now pays for a 192-byte Opt through the State interface. A comment right after the destination skips the speculative option pass, without which the long comment was 44% slower. All Benchmark_Parser_Next* stay at 0 allocs/op, and the VM check path keeps its AllocsPerRun == 0 guard.

Validation

Tests were rewritten first and failed: every InlineComment assertion became a full ReduceState with the OptComment option, plus new cases for CRLF, not //, comments inside groups, a state rejecting the comment, hooks next to comments, and Test_VM_Check_CommentOption for the verdicts. make test (race) and make lint pass.