Speedup DenseOps compilation

[pinzhenx]

This PR follows the same approach as what pytorch has done for VariableType.cpp, to split the all-in-one DenseOps.cpp into 8 shards, called DenseOps_0.cpp ... DenseOps_7.cpp.

We've removed RegisterIpexDenseOPs() and all occurrences of _initialize_aten_bindings(), since we could register ops in global space while keep its internal linkage. This way, we don't have to worry about multiple calls to the RegisterIpexDenseOPs.

Before:

void RegisterIpexDenseOPs() {
  static auto dispatch = torch::RegisterOperators()
    .op(torch::RegisterOperators::options().schema("aten::__and__.Scalar(Tensor self, Scalar other) -> Tensor")
      .impl_unboxedOnlyKernel<at::Tensor(const at::Tensor &, at::Scalar), &AtenIpexCPUDefault::__and__>(at::DispatchKey::DPCPPTensorId)
      .aliasAnalysis(c10::AliasAnalysisKind::FROM_SCHEMA));
}

After:

namespace {
  static auto dispatch = torch::RegisterOperators()
    .op(torch::RegisterOperators::options().schema("aten::__and__.Tensor(Tensor self, Tensor other) -> Tensor")
      .impl_unboxedOnlyKernel<at::Tensor(const at::Tensor &, const at::Tensor &), &AtenIpexCPUDefault::__and__>(at::DispatchKey::DPCPPTensorId)
      .aliasAnalysis(c10::AliasAnalysisKind::FROM_SCHEMA));
}

After checkout this branch, you have to run python setup.py clean first.

[pinzhenx]

@EikanWang

[hongzhen1]

better integrate these two packages into one, so that end user only needs to import one package.