OneHot missing kernel for float data type with int32 indices

[fdwr]

Describe the bug
Lotus's OneHot implementation registers as many as 6 permutations, yet it's missing the most intuitive combination, float values with integer indices.

onnxruntime\core\providers\cpu\tensor\onehot.cc

#define REG_ONE_HOT_OP(index_type, out_type, depth_type)
  REG_TYPED_ONE_HOT_OP(index_type##_##out_type##_##depth_type, in_type, out_type, depth_type)

REG_ONE_HOT_OP(int64_t, int64_t, int64_t);
REG_ONE_HOT_OP(float, int64_t, int64_t);
REG_ONE_HOT_OP(int64_t, string, int64_t);
REG_ONE_HOT_OP(float, string, int64_t);
REG_ONE_HOT_OP(float, float, float);      // added this to satisfy onnx model tests
REG_ONE_HOT_OP(int64_t, int32_t, float);  // added this to satisfy onnx model tests

There is an entry with float indices and int64_t output values (which feels like a swapped typo, as indices are typically integers rather than floats, and int64_t in an output tensor times the depth is huge); and there is an entry for int32_t outputs with float depth, which is weird (added just for ONNX model tests, but doesn't seem like the most realistic case). I propose adding either of these:

REG_ONE_HOT_OP(int64_t, float, int64_t); // integer indices and depth, float value
REG_ONE_HOT_OP(uint32_t, float, uint32_t); // integer indices and depth, float value

Supporting int32 (or uint32) makes sense to keep memory usage down, given a single axis will not be larger than 4 billion elements (even if a tensor as a whole is larger than 4 billion elements, a single axis shouldn't be). DML supports int32/int64 with float.

System information

• Windows 10 OS WinML 18931.1000
• ONNX Runtime installed from (source or binary): source
• ONNX Runtime version: 1.5
• Python version: NA
• Visual Studio version (if applicable): Visual Studio 2017
• GCC/Compiler version (if compiling from source): NA
• CUDA/cuDNN version: NA
• GPU model and memory: NA

To Reproduce

{
  "op_type": "OneHot",
  "indices": [1, 0, 3],
  "depth": [4],
  "values": [1, 2],
  "output": [[1, 2, 1],
             [2, 1, 1],
             [1, 1, 1],
             [1, 1, 2]],
  "axis": 0,
  "T1": "int64", // indices
  "T2": "int64", // depth
  "T3": "float32" // value
},

Expected behavior
Support integer indices with floating data type values.

[hariharans29]

Closed via #1317