feat(decompiler): Rec 31 #31-3 RAII Stage 2B — xml_parse global_scan unique_ptr

[CryptoJones]

Stacked on #51 — base is feat/rec-31-raii-stage2b-xml-lvalue, not master. GitHub will rebase to master automatically when #51 lands. The diff visible here is only the global_scan change; the lvalue changes are part of #51's scope.

Converts xml_parse()'s pairing of global_scan = new XmlScan(i) / delete global_scan; to a scope-bound unique_ptr that holds ownership across the yyparse call:

auto scan = make_unique<XmlScan>(i);
global_scan = scan.get();    // raw observer pointer for grammar actions
...
int4 res = yyparse();
...
global_scan = (XmlScan *)0;  // null observer before scan goes out of scope
return res;

global_scan stays a static XmlScan * (raw observer) because it's accessed from yyparse / yylex / grammar semantic actions — those see it through the symbol declared at file scope, and changing it to unique_ptr would require changing every accessor. The observer pattern keeps the surface area minimal: only xml_parse owns; everyone else just borrows during the parse.

The explicit global_scan = nullptr before return is defensive — if anything ever tries to dereference global_scan after xml_parse exits, crash on null is far better than use-after-free on dangling.

After this PR, the only raw news remaining in the xml.y / xml.cc epilogue are:

• line 538 new Element(cur) — parse-tree node allocation, owned by parent's children list; needs Element class field refactor.
• line 624 new Document() — returned across xml_tree() API to XmlDecode; needs xml.hh + XmlDecode coordination.

Both are separate PRs. The bison semantic-action raw news (xml.y:150, 153, 198, 200, 208) remain blocked on %union redesign.

Proudly Made in Nebraska. Go Big Red! 🌽 https://xkcd.com/2347/