[BUG] Incorrect type check in engine.py for CPU training

[learning-chip]

Describe the bug

The type check inside function split_half_float_double_sparse of runtime/engine.py does not recognize CPU tensors. Simply because it is checking for string "torch.cpu.FloatTensor" while the CPU tensor type is "torch.FloatTensor" (no "cpu")

https://github.com/microsoft/DeepSpeed/blob/fc9e1ee00e673b2ce2a433c4e34c5440e76c9f3e/deepspeed/runtime/engine.py#L115-L124

To Reproduce

Enable Intel CPU backend of deepspeed:

pip install torch==1.13.1+cpu
pip install deepspeed==0.9.5
pip install intel_extension_for_pytorch==1.13+cpu -f https://developer.intel.com/ipex-whl-stable-cpu
pip install oneccl_bind_pt==1.13+cpu -f https://developer.intel.com/ipex-whl-stable-cpu
# also need to build oneCCL itself

Modify small_model_debugging/test_model.py to work on CPU:

• set "torch_adam": True in config_dict to skip op builder, and set "enabled": False for "fp16"
• set dtype=torch.float for train_data to avoid half precision

Running the script leads to error AssertionError: attempting to reduce an unsupported grad type: torch.FloatTensor. Remove the assertion (line 123~124 in DS 0.9.5) then it trains fine.