Switch from pybind11 to nanobind

[inducer]

Surprisingly, this is mostly working now. :)

To do/Needs:

• [x] Pass location of numpy headers to CMake (found via CMake now)
• [x] Pass OpenCL config info to CMake (found via CMake now)
• [x] Actually fix the wrapper :slightly_smiling_face:
  • [x] pycl test_wrapper.py 'test_mempool(cl._csc)' crashes
  • [x] PYOPENCL_TEST=port python test_array.py "test_mem_pool_with_arrays(cl._csc)" crashes
  • [x] Doesn't make it through the meshmode test suite
• [x] https://github.com/wjakob/nanobind/pull/38
• [x] https://github.com/wjakob/nanobind/pull/23
• [x] https://github.com/wjakob/nanobind/pull/24
• [x] See if the _enqueue trampoline can be avoided (doesn't matter for loopy enqueues, that goes straight to enqueue_nd_range_kernel)
• [x] https://github.com/wjakob/nanobind/pull/71
• [x] Expose SVM-as-buffer
• [x] Investigate reference leaks after tests: they've disappeared.
• [x] Wait for nanobind 2, for for_setter.
• [x] Pass CI
• [x] Check loopy CI. Currently, it requires weak references to cl.Context, which nanobind doesn't currently allow. (Edit: Now does allow.)
• [x] Work around https://github.com/wjakob/nanobind/pull/72 (This seems to have been resolved in the meantime.)

[inducer]

The infinite recursions appear to be https://github.com/wjakob/nanobind/pull/72.

[inducer]

@matthiasdiener FYI this is happening I think. I'll probably roll a release once it's in.