Added pow to autocast policy and unit test

[avizon-aws]

Pow is a non linear operation and must be done in higher precision i.e. FP32.

Before this CR:

import torch
import torch_xla.core.xla_model as xm
input=torch.rand((5,5), dtype=torch.bfloat16).to(xm.xla_device())
with torch.autocast("xla"):
    output=input.pow(2) // this should happen in FP32
xm.mark_step()

HloModule SyncTensorsGraph.6, entry_computation_layout={(bf16[5,5]{1,0})->(bf16[5,5]{1,0})}

ENTRY %SyncTensorsGraph.6 (p0.2: bf16[5,5]) -> (bf16[5,5]) {
  %p0.2 = bf16[5,5]{1,0} parameter(0), frontend_attributes={neff_input_names="input0"}, metadata={op_type="xla__device_data" op_name="xla__device_data" source_file="/shared/avizon/avi_trial_exps/test.py" source_line=4}
  %multiply = bf16[5,5]{1,0} multiply(bf16[5,5]{1,0} %p0.2, bf16[5,5]{1,0} %p0.2), metadata={op_type="aten__pow" op_name="aten__pow" source_file="/shared/avizon/avi_trial_exps/test.py" source_line=7}
  ROOT %tuple.5 = (bf16[5,5]{1,0}) tuple(bf16[5,5]{1,0} %multiply), frontend_attributes={neff_output_names="output0"}
}

//Its happening in BF16, which is wrong

after the CR:

HloModule SyncTensorsGraph.7, entry_computation_layout={(bf16[5,5]{1,0})->(f32[5,5]{1,0})}

ENTRY %SyncTensorsGraph.7 (p0.2: bf16[5,5]) -> (f32[5,5]) {
  %p0.2 = bf16[5,5]{1,0} parameter(0), frontend_attributes={neff_input_names="input0"}, metadata={op_type="xla__device_data" op_name="xla__device_data" source_file="/shared/avizon/avi_trial_exps/test.py" source_line=4}
  %convert.3 = f32[5,5]{1,0} convert(bf16[5,5]{1,0} %p0.2), metadata={op_type="xla__cast" op_name="xla__cast" source_file="/shared/avizon/avi_trial_exps/test.py" source_line=7}
  %multiply = f32[5,5]{1,0} multiply(f32[5,5]{1,0} %convert.3, f32[5,5]{1,0} %convert.3), metadata={op_type="aten__pow" op_name="aten__pow" source_file="/shared/avizon/avi_trial_exps/test.py" source_line=7}
  ROOT %tuple.6 = (f32[5,5]{1,0}) tuple(f32[5,5]{1,0} %multiply), frontend_attributes={neff_output_names="output0"}
}

//The inputs have been upcasted