Metadata agnostic user computation hash

[rpsilva-aws]

Cherry pick of #8550

This fixes #8538.

Test context:

• input_scope_0

HloModule IrToHlo.16, entry_computation_layout={(s64[])->(f32[])}

%computation.6 (p0.7: f32[], p1.8: f32[]) -> (f32[]) {
  %p0.7 = f32[] parameter(0)
  %p1.8 = f32[] parameter(1)
  %maximum.9 = f32[] maximum(f32[] %p0.7, f32[] %p1.8)
  %minimum.10 = f32[] minimum(f32[] %p0.7, f32[] %p1.8)
  %subtract.11 = f32[] subtract(f32[] %maximum.9, f32[] %minimum.10)
  ROOT %tuple.12 = (f32[]) tuple(f32[] %subtract.11)
}

ENTRY %IrToHlo.16 (p0.1: s64[]) -> (f32[]) {
  %p0.1 = s64[] parameter(0), metadata={op_type="xla__device_data" op_name="xla__device_data" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=43}
  %convert.4 = f32[] convert(s64[] %p0.1), metadata={op_type="aten__sin" op_name="aten__sin" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=32}
  %sine.5 = f32[] sine(f32[] %convert.4), metadata={op_type="aten__sin" op_name="aten__sin" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=32}
  %convert.2 = f32[] convert(s64[] %p0.1), metadata={op_type="aten__cos" op_name="aten__cos" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=33}
  %cosine.3 = f32[] cosine(f32[] %convert.2), metadata={op_type="aten__cos" op_name="aten__cos" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=33}
  %call.13 = (f32[]) call(f32[] %sine.5, f32[] %cosine.3), to_apply=%computation.6, metadata={op_type="xla__computation" op_name="xla__computation" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=50}
  %get-tuple-element.14 = f32[] get-tuple-element((f32[]) %call.13), index=0, metadata={op_type="xla__computation" op_name="xla__computation" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=50}
  ROOT %tuple.15 = (f32[]) tuple(f32[] %get-tuple-element.14)
}

• input_scope_1

HloModule IrToHlo.16, entry_computation_layout={(s64[])->(f32[])}

%computation.6 (p0.7: f32[], p1.8: f32[]) -> (f32[]) {
  %p0.7 = f32[] parameter(0)
  %p1.8 = f32[] parameter(1)
  %maximum.9 = f32[] maximum(f32[] %p0.7, f32[] %p1.8)
  %minimum.10 = f32[] minimum(f32[] %p0.7, f32[] %p1.8)
  %subtract.11 = f32[] subtract(f32[] %maximum.9, f32[] %minimum.10)
  ROOT %tuple.12 = (f32[]) tuple(f32[] %subtract.11)
}

ENTRY %IrToHlo.16 (p0.1: s64[]) -> (f32[]) {
  %p0.1 = s64[] parameter(0), metadata={op_type="xla__device_data" op_name="xla__device_data" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=38}
  %convert.4 = f32[] convert(s64[] %p0.1), metadata={op_type="aten__sin" op_name="aten__sin" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=38}
  %sine.5 = f32[] sine(f32[] %convert.4), metadata={op_type="aten__sin" op_name="aten__sin" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=38}
  %convert.2 = f32[] convert(s64[] %p0.1), metadata={op_type="aten__cos" op_name="aten__cos" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=39}
  %cosine.3 = f32[] cosine(f32[] %convert.2), metadata={op_type="aten__cos" op_name="aten__cos" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=39}
  %call.13 = (f32[]) call(f32[] %sine.5, f32[] %cosine.3), to_apply=%computation.6, metadata={op_type="xla__computation" op_name="xla__computation" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=50}
  %get-tuple-element.14 = f32[] get-tuple-element((f32[]) %call.13), index=0, metadata={op_type="xla__computation" op_name="xla__computation" source_file="/ansible/pytorch/xla/test_user_computation_debug_cache.py" source_line=50}
  ROOT %tuple.15 = (f32[]) tuple(f32[] %get-tuple-element.14)
}