Improved efficiency of pybind11 classes and type casters.

Type casters in pybind11 generated with PYBIND11_TYPE_CASTER macro rely on default constructor of C++ type. The default constructed value is then overwritten by load method. C++ types such as sycl::device and sycl::queue do non-trivial amount of work in default constructor. This work was ranking high in certain workloads (such as example/pybind11/onemkl_gemv/sycl_timing_solver.py).

Same applies to auto-generated type-casters for classes dpctl::memory::usm_memory and dpctl::tensor::usm_ndarray.

This PR introduces DPCTL_TYPE_CASTER macro that defines unique_ptr to the C++ type, rather than the type itself. This avoids superfluous call to default constructor altogether.

Instead of redefining pyobject_caster automatically generated for dpctl::tensor::usm_ndarray and dpctl::memory::usm_memory a singleton class dpctl::detail::dpctl_api is added that creates objects used by default constructors of these classes.

• Have you provided a meaningful PR description?
• Have you added a test, reproducer or referred to an issue with a reproducer?
• Have you tested your changes locally for CPU and GPU devices?
• Have you made sure that new changes do not introduce compiler warnings?
• If this PR is a work in progress, are you filing the PR as a draft?