[CUDA] Work around a bug in rint() caused by a broken implementation provided by CUDA.

[tra]

Summary

Normally math functions are forwarded to nv_* counterparts provided by CUDA's
libdevice bitcode. However, nv_rint*() functions there have a bug -- they use
round() which rounds *up* instead of rounding towards the nearest integer, so we
end up with rint(2.5f) producing 3.0 instead of expected 2.0. The broken bitcode
is not actually used by NVCC itself, which has both a work-around in CUDA
headers and, in recent versions, uses correct implementations in NVCC's built-ins.

This patch implements equivalent workaround and directs rint/rintf to
__builtin_rint/rintf that produce correct results.

[jlebar]

LGTM, and can we write a test in the test-suite?

[tra]

Also fixed the same bug in nearbyint().

clang/lib/Headers/__clang_cuda_math.h

 __DEVICE__ long lrintf(float __a) { return __float2int_rn(__a); }
 __DEVICE__ long lround(double __a) { return round(__a); }
 __DEVICE__ long lroundf(float __a) { return roundf(__a); }
 #endif
 __DEVICE__ int max(int __a, int __b) { return __nv_max(__a, __b); }
 __DEVICE__ int min(int __a, int __b) { return __nv_min(__a, __b); }
 __DEVICE__ double modf(double __a, double *__b) { return __nv_modf(__a, __b); }
 __DEVICE__ float modff(float __a, float *__b) { return __nv_modff(__a, __b); }
-__DEVICE__ double nearbyint(double __a) { return __nv_nearbyint(__a); }
-__DEVICE__ float nearbyintf(float __a) { return __nv_nearbyintf(__a); }
+__DEVICE__ double nearbyint(double __a) { return __builtin_nearbyint(__a); }
+__DEVICE__ float nearbyintf(float __a) { return __builtin_nearbyintf(__a); }
 __DEVICE__ double nextafter(double __a, double __b) {
   return __nv_nextafter(__a, __b);
 }
 __DEVICE__ float nextafterf(float __a, float __b) {
   return __nv_nextafterf(__a, __b);
 }
 __DEVICE__ double norm(int __dim, const double *__t) {
   return __nv_norm(__dim, __t);
__DEVICE__ float remquof(float __a, float __b, int *__c) {
   return __nv_remquof(__a, __b, __c);
 }
 __DEVICE__ double rhypot(double __a, double __b) {
   return __nv_rhypot(__a, __b);
 }
 __DEVICE__ float rhypotf(float __a, float __b) {
   return __nv_rhypotf(__a, __b);
 }
-__DEVICE__ double rint(double __a) { return __nv_rint(__a); }
-__DEVICE__ float rintf(float __a) { return __nv_rintf(__a); }
+// __nv_rint* in libdevice is buggy and produces incorrect results.
+__DEVICE__ double rint(double __a) { return __builtin_rint(__a); }
+__DEVICE__ float rintf(float __a) { return __builtin_rintf(__a); }
 __DEVICE__ double rnorm(int __a, const double *__b) {
   return __nv_rnorm(__a, __b);
 }
 __DEVICE__ double rnorm3d(double __a, double __b, double __c) {
   return __nv_rnorm3d(__a, __b, __c);
 }
 __DEVICE__ float rnorm3df(float __a, float __b, float __c) {
   return __nv_rnorm3df(__a, __b, __c);