Switch round intrinsic to use round call across all platforms (instead of nearbyint on some)

[vksnk]

From #6406.

Currently, round intrinsic in CodeGen_LLVM is implemented via nearbyint() call. However, on some of the platforms it's implemented via round call. We should make this consistent and implement it via round(). Also, we may want to introduce a separate intrinsic for nearby.

[abadams]

Note

Automated triage: this issue was flagged by Claude as likely safe to close, for the reason below. The reasoning was checked against the current source, and a second pass tried and failed to refute it, but no human verified this specific text. If the reason is wrong, please reopen the issue and say so — some fraction of these are expected to be wrong.

Reason (CLOSE_FIXED): The issue was about Call::round having platform-dependent semantics (nearbyint on some backends, round on others). That inconsistency was resolved by standardizing on round-to-nearest-ties-to-even everywhere, with a portable software fallback, and documenting it.

Evidence: src/CodeGen_LLVM.cpp:3102-3107 emits llvm.roundeven and falls back to lower_round_to_nearest_ties_to_even(); src/CodeGen_C.cpp:1624-1628 lowers the same way; src/CodeGen_ARM.cpp:436-441 and src/CodeGen_X86.cpp:111-114 map "round" to roundeven/roundps; documented at src/IROperator.h:1042-1047; commits d5a81186a (#7012 "Make Halide::round behave as documented") and d3a8638bd (#8926); covered by test/correctness/round.cpp.

Resolved by #7012 and #8926, though in the opposite direction from the title: Call::round is now round-to-nearest-ties-to-even on every backend (llvm.roundeven where available, lower_round_to_nearest_ties_to_even() otherwise, including CodeGen_C), and IROperator.h documents that contract. Closing; if a separate away-from-zero or nearbyint intrinsic is wanted, that is a fresh feature request.