Allow CUDA source inputs compiled to LTOIR, and enable pynvjitlinker to link inputs that contains LTOIR

[isVoid]

This PR supercedes #60 due to write permission issue.

[isVoid]

I'm not able to reproduce this segfault on my v100 machine:

test_namedunituple (numba.cuda.tests.cudapy.test_array_args.TestCudaArrayArg.test_namedunituple) ... Fatal Python error: Segmentation fault
...

Extension modules: numpy._core._multiarray_umath, numpy._core._multiarray_tests, numpy.linalg._umath_linalg, numba.core.typeconv._typeconv, numpy.random._common, numpy.random.bit_generator, numpy.random._bounded_integers, numpy.random._mt19937, numpy.random.mtrand, numpy.random._philox, numpy.random._pcg64, numpy.random._sfc64, numpy.random._generator, numba._helperlib, numba._dynfunc, numba._dispatcher, numba.core.runtime._nrt_python, numba.np.ufunc._internal, numba.experimental.jitclass._box, numba.mviewbuf, pynvjitlink._nvjitlinklib, numba.types.itertools (total: 22)
ci/test_conda_pynvjitlink.sh: line 72:  2238 Segmentation fault      (core dumped) ENABLE_PYNVJITLINK=1 NUMBA_CUDA_TEST_BIN_DIR=$NUMBA_CUDA_TEST_BIN_DIR python -m numba.runtests numba.cuda.tests -v
/__w/numba-cuda/numba-cuda

[isVoid]

In afcce87 I added an additional flag ignore_nonlto to the linker so that only LTO-able objects are added when the flag is enabled. This allows the driver is in the correct state when -ptx flag is set. And this is also the desired behavior since Numba now dumps the optimized PTX only for the portion that are LTO-abled added to the linker, and raise warning for any source that aren't optimizable.

[isVoid]

A subtle case here is that lto=True is only enabled for cuda>12.0 and is tested so. Because this feature depends on pynvjitlink, which is only tested in CTK12.5 environment.

[isVoid]

There's a nuance here where fatbin object can also only contains LTOIR. Though that requires building additional bindings for cuobjdump and outputting the result.

[isVoid]

In bac11f6 I made a new created a new fatbin object that contains both LTOIR and SASS and feed into the linker for testing. I added a helper function that invokes cuobjdump to extract object types from the fatbin / object. We allow and fatbin / object that contains LTOIR to be passed into the linker.

[gmarkall]

After a merge from main, this passes all tests for me locally on Windows.

[gmarkall]

Suggested change

	When `ignore_nonlto` is set to true, do not add code that are will not
	When `ignore_nonlto` is set to true, do not add code that will not

[gmarkall]

This tests OK for me locally on Windows. I've pushed changes to the Windows test binary build script that mirror the Makefile changes.

I think this is good to merge, pending:

• Completion of CI
• @isVoid checking integration with Numbast again locally, with the recent changes.