This is to avoid multiple definitions compilation error when using
multiple translation units.

…efinition of' compilation error messages

Moved implementation of kernels out to dedicated header files.

…d it

Use -O1 when compiling tensor_py for now to work around suspected issue
with loading of C-API functions.

…es`.

The class is already singleton. Instead create a local variable at
each use site. This local variable is going to be a constant reference
to the singleton.

…ll *.cpp

[cleanup/tensor, part 2] Modularized tests for contiguity, retrieval of PyUSMArrayObject* and removed use of a global variable.

The class implements a singleton pattern that populates global
variables exported by Cython CAPI functions via call to import_dpctl()
and then populates internal struct members with pointers populated
by Cython.

This loads both Cython 'public api' functions, types, and integer values.

dpctl::tensor::usm_ndarray and dpctl::memory::usm_memory classes, as
well as type-casters for dpctl.SyclQueue <-> sycl::queue,
dpctl.SyclDevice <-> sycl::device, etc. use dpctl_capi.

The singleton patterns helps avoid static initialization order fiasco
problem when using multiple translation units.

Single dpctl_capi loads typenum values, the
dpctl::tensor::detail::usm_ndarray_types struct no longer needs to be
a singleton, and becomes an ordinary struct with a static method.

tensor_py.cpp was changed to reflect that.

[cleanup/tensor, step 3] Implement dpctl_capi loader in dpctl4pybind11 header file

This is avoid using py::value_error and py::ssize_t and py::type_error
in there explicitly.

Use in tensor_py.cpp adjusted

[cleanup/tensor, part 4] Made contract_iter and contract_iter2 functions templated to avoid needing to include pybind11 headers

…ed file

Also moved routine for simplification of iteration space into dedicated file