[Flang][OpenMP] Support conditional lastprivate on host (#200086)

This patch lowers lastprivate(conditional:) on the host by leveraging
the existing user-defined reduction (UDR) infrastructure. A packed
struct is created where each thread tracks, for every lastprivate
variable, both the candidate value and the canonical iteration index of
its last update. The reduction combiner selects the value from the
sequentially later iteration (for do loops) or from the lexically later
section (for sections).

The implementation locates the enclosing omp.parallel and places the
shared struct before it so that all threads in the team reduce into the
same storage. For orphaned worksharing constructs — where no enclosing
parallel is visible at compile time — a module-scope global of the
struct type is used instead. This is correct for a single level of
parallelism, but concurrent nested teams executing the same orphaned
construct would race on the shared global. This limitation mirrors the
current Clang behavior, which also uses a single global and does not
support nested parallelism for conditional lastprivate.

Assisted-by : Claude Opus 4.6