Workarounds to enable DeepSpeed transformer inference on ROCm

[rraminen]

These changes are required to resolve the following errors when running bloom workload:

1. /opt/conda/lib/python3.8/site-packages/deepspeed/ops/csrc/includes/conversion_utils_hip.h:270:12: error: use of undeclared identifier '__double2half'; did you mean '__double2hiint'?
   return __double2half(val);
          ^~~~~~~~~~~~~
          __double2hiint
   /opt/rocm-5.4.0/include/hip/amd_detail/amd_device_functions.h:440:30: note: '__double2hiint' declared here
   __device__ static inline int __double2hiint(double x) {
                            ^

2. 
   /opt/conda/lib/python3.8/site-packages/deepspeed/ops/csrc/transformer/inference/csrc/apply_rotary_pos_emb.hip:195:48: error: use of undeclared identifier '__shfl_sync'
           auto q_rot_tmp = lane < half_dim ? __shfl_sync(mask[lane], q_rot, lane + half_dim)

3. 

/opt/conda/lib/python3.8/site-packages/deepspeed/ops/csrc/includes/reduction_utils_hip.h:278:43: error: excess elements in struct initializer constexpr __half2_raw zero = {0x0000, 0x0000}; ^~

[rraminen]

Use https://github.com/ROCmSoftwarePlatform/DeepSpeed/tree/transformer_inference branch which contains the changes mentioned in this PR

Closing this PR as we have the below two merged

1. https://github.com/ROCmSoftwarePlatform/DeepSpeed/pull/59
2. https://github.com/ROCmSoftwarePlatform/DeepSpeed/pull/60