[clang] [llvm] [CUDA][NVPTX] Add dp4a/dp2a builtins; define CUDA-13.4's __nv_idp*_impl externs (PR #233434)

[Ilya Biryukov via cfe-commits]

================
@@ -203,6 +203,30 @@ let Attributes = [NoThrow] in {
   def __nvvm_pm_event_mask : NVPTXBuiltin<"void(_Constant unsigned short)">;
 }
 
+// Integer dot product (dp4a / dp2a).
+//
+// __nvvm_idp4a_<a>_<b>(a, b, c): four-way byte dot product of a and b plus c.
+// __nvvm_idp2a_{lo,hi}_<a>_<b>(a, b, c): two-way dot product of the 16-bit
+// halves of a with the low/high byte pair of b, plus c.
+// <a>, <b> are the signedness ('s' or 'u') of the packed elements of a and b.
+// The result is signed unless both inputs are unsigned.
----------------
ilya-biryukov wrote:

Removed it from here.

https://github.com/llvm/llvm-project/pull/233434