Remove UB type punning from c10/util/floating_point_utils.h (pytorch#140567)

Accessing the inactive member of a union is undefined behavior. Fortunately, we have c10::bit_cast.

Differential Revision: [D65888680](https://our.internmc.facebook.com/intern/diff/D65888680/)
Pull Request resolved: pytorch#140567
Approved by: https://github.com/Skylion007, https://github.com/malfet
ghstack dependencies: pytorch#140564, pytorch#140565, pytorch#140566

Author: swolchok

c10/util/floating_point_utils.h

@@ -1,39 +1,32 @@
 #pragma once
 
 #include <c10/macros/Macros.h>
+#include <c10/util/bit_cast.h>
 #include <cstdint>
 
 namespace c10::detail {
 
 C10_HOST_DEVICE inline float fp32_from_bits(uint32_t w) {
 #if defined(__OPENCL_VERSION__)
   return as_float(w);
-#elif defined(__CUDA_ARCH__)
+#elif defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__)
   return __uint_as_float((unsigned int)w);
 #elif defined(__INTEL_COMPILER)
   return _castu32_f32(w);
 #else
-  union {
-    uint32_t as_bits;
-    float as_value;
-  } fp32 = {w};
-  return fp32.as_value;
+  return c10::bit_cast<float>(w);
 #endif
 }
 
 C10_HOST_DEVICE inline uint32_t fp32_to_bits(float f) {
 #if defined(__OPENCL_VERSION__)
   return as_uint(f);
-#elif defined(__CUDA_ARCH__)
+#elif defined(__CUDA_ARCH__) || defined(__HIP_DEVICE_COMPILE__)
   return (uint32_t)__float_as_uint(f);
 #elif defined(__INTEL_COMPILER)
   return _castf32_u32(f);
 #else
-  union {
-    float as_value;
-    uint32_t as_bits;
-  } fp32 = {f};
-  return fp32.as_bits;
+  return c10::bit_cast<uint32_t>(f);
 #endif
 }