Clean up WeightRow in preparation for optimizer state offloading

Summary:
X-link: facebookresearch/FBGEMM#1109

- Clean up `WeightRow` implementation in preparation for optimizer state offloading
- Add documentation for the class

Differential Revision: D73473546

Author: q10

fbgemm_gpu/codegen/training/forward/embedding_forward_split_meta_template.cpp

@@ -33,7 +33,7 @@
 using namespace fbgemm_gpu;
 using Tensor = at::Tensor;
 
-[[maybe_unused]] static constexpr float kINT8QparamsBytes = 8;
+[[maybe_unused]] static constexpr int32_t kINT8QparamsBytes = 8;
 
 ////////////////////////////////////////////////////////////////////////////////
 // Kernel Definitions

fbgemm_gpu/include/fbgemm_gpu/utils/cuda_prelude.cuh

@@ -85,7 +85,7 @@ static constexpr float kQParamEps = 1e-8f;
 will be stored at the end of each row in FP32 formats, appending a total of
 8 bytes to each row.
 */
-static constexpr float kINT8QparamsBytes = 8;
+static constexpr int32_t kINT8QparamsBytes = 8;
 
 template <typename T>
 DEVICE_INLINE T shfl_xor(