[INFRA] Import NVIDIA/CCCL upstream as optimization reference library

CCCL (CUDA C++ Core Libraries) provides:
- CUB: device/block/warp-level GPU primitives (reduce, scan, sort, topk)
- Thrust: high-level parallel algorithms (transform_reduce, sort, scan)
- libcudacxx: CUDA C++ standard library (atomics, barriers, memory)
- cudax: experimental features (memory resources, allocators)
- Tuning policies: per-SM hardware-specific algorithm parameters

Competition optimization vectors mapped to CCCL:
- Output TPS (83% weight): warp_reduce, block_reduce, device_topk
- Input TPS (14% weight): device_scan, block_load, prefetch
- Cache TPS (3% weight): prefix caching strategy patterns
- Memory (0.9 util): pooled/cached/buddy allocators

Source: https://github.com/NVIDIA/cccl (shallow clone, HEAD only)
License: Apache-2.0

cccl_upstream/libcudacxx/test/support/hexfloat.h

@@ -0,0 +1,57 @@
+//===----------------------------------------------------------------------===//
+//
+// Part of the LLVM Project, under the Apache License v2.0 with LLVM Exceptions.
+// See https://llvm.org/LICENSE.txt for license information.
+// SPDX-License-Identifier: Apache-2.0 WITH LLVM-exception
+//
+//===----------------------------------------------------------------------===//
+
+// Define a hexfloat literal emulator since we can't depend on being able to
+//   for hexfloat literals
+
+// 0x10.F5p-10 == hexfloat<double>(0x10, 0xF5, -10)
+
+#ifndef HEXFLOAT_H
+#define HEXFLOAT_H
+
+#include <climits>
+#include <cmath>
+
+template <class T>
+class hexfloat
+{
+  T value_;
+
+  static int CountLeadingZeros(unsigned long long n)
+  {
+    const std::size_t Digits        = sizeof(unsigned long long) * CHAR_BIT;
+    const unsigned long long TopBit = 1ull << (Digits - 1);
+    if (n == 0)
+    {
+      return Digits;
+    }
+    int LeadingZeros = 0;
+    while ((n & TopBit) == 0)
+    {
+      ++LeadingZeros;
+      n <<= 1;
+    }
+    return LeadingZeros;
+  }
+
+public:
+  hexfloat(long long m1, unsigned long long m0, int exp)
+  {
+    const std::size_t Digits = sizeof(unsigned long long) * CHAR_BIT;
+    int s                    = m1 < 0 ? -1 : 1;
+    int exp2                 = -static_cast<int>(Digits - CountLeadingZeros(m0) / 4 * 4);
+    value_                   = std::ldexp(m1 + s * std::ldexp(T(m0), exp2), exp);
+  }
+
+  operator T() const
+  {
+    return value_;
+  }
+};
+
+#endif