Improve rule stacking

Current rule stacking actually runs the preprocessor as well as the rule engine twice, which can't help performance. I'm thinking we could instead concatenate the two rules before PP and only run things once although that will involve less trivial things for throwing things around (eg. if the second rule has reject-flags they must be moved to just after the first rule's reject flags (if any) and/or remove dupes of such.

Example, rule1 ^[A-Z] and rule2 Az"[0-9][0-9]"

Current flow (first entry from PP from respective rule):

"word" -> apply ^A -> "Aword" -> apply Az"00" -> "Aword00"

New flow, single rule, PP ^[A-Z] Az"[0-9][0-9]"

"word" -> apply ^A Az"00" -> "Aword00"

That was easy. Now something less trivial:

Rule1 -p -c (?a 2 (?a c 1 [cl] and rule2 -c /?v V Q

Current flow (single mode, 2 words)

"john"+"smith" -> -p -c (?a 2 (?a c 1 c -> "JohnSmith" -> -c /?v V Q -> "JoHNSMiTH"

New flow, concat -p -c (?a 2 (?a c 1 [cl].+ M.-c /?v V Q (M always added between them, and in this case also + because we have 1 and/or 2 in the first rule) and then dropping the second -c results in -p -c (?a 2 (?a c 1 [cl] + M /?v V Q

"john"+"smith" -> -p -c (?a 2 (?a c 1 c + M /?v V Q -> "JoHNSMiTH"

A drawback with doing/fixing this is that current --rules-stack run first word through all rules before going on to next word (normal rules do the opposite). This is sometimes wanted.

Start with the --rules-stack path and the preprocessor and rule-engine flow described in the issue. Compare the current two-pass behavior with the proposed concatenated-rule examples, including reject flags, separators, and the different word-order behavior. Done means stacked rules can be processed once while preserving the intended output and handling those edge cases.