ICU 72 - prevent stack overflow in collator

[ilonatommy]

Problem this PR is fixing

See: dotnet/runtime#121202 (comment).

Recursive canonical-closure expansion in ICU’s canonical iterator and collation importer could revisit the same substring/import chain indefinitely. With deeply nested or cyclic decompositions/imports, the iterator would overflow the stack and [import …] would spin forever (or until the host stack failed).

Fixes

• Canonical iterator caniter.cpp: added a cycle guard. Every segment encountered during the recursive walk is tracked in a “visited” set collectEquivalentsRecursive so we stop recursing once we see the same canonical form again. Prevents runaway recursion and keeps closure generation finite.
• Collation rule parser collationruleparser.cpp introduced importRecursionGuard, a hash table that records the [import locale/type] pairs currently being processed. If an import re-enters one of those pairs, we break out instead of chasing the loop, stopping the stack overflow. Normalized the locale/type strings before storing them, and updated the cleanup so the guard is cleared as we unwind.

Together these changes keep both canonical closure expansion and rule imports from recursing indefinitely while touching only the minimal code paths needed for the stack overflow fix.

Filters were restored to the state from before #686 and prebuilts were updated accordingly.