Claim the bundle's vendor libraries, and drop the rocSPARSE version probe

[luraess]

EDIT: rolled-back most of #1001

In artifact mode on a host with another ROCm on LD_LIBRARY_PATH (here a CSCS uenv), versioninfo() reported rocSPARSE as err and the core test asserting the probe returns a version failed. The cause is not rocSPARSE.

PRELOAD_LIBRARIES claimed the runtime closure but no vendor library, and AMDGPU claims those lazily on the first ccall, so whichever is touched first wins. TheRock's libraries carry DT_RUNPATH, which LD_LIBRARY_PATH outranks, so touching rocSPARSE first pulled its dependencies from the environment. Measured in artifact mode after rocSPARSE.version():

rocsparse -> <artifact>/lib/librocsparse.so
rocblas   -> /user-environment/env/default/lib/librocblas.so.5
roctx     -> /user-environment/env/default/lib/libroctx64.so.4
foreign_libraries() = String[]

FOREIGN_LIBRARY_NAMES listed only runtime libraries, so versioninfo() printed "Foreign ROCm libraries loaded: none" while two rocBLAS versions were live in the process.

The isolated probe from #1051 hit the same thing: a bare child has claimed nothing at all, so it resolved the bundle's librocsparse against uenv HIP 7.2.3 and rocsparse_create_handle segfaulted, which is what surfaced as err.

Changes

• PRELOAD_LIBRARIES also claims libamdhip64, libroctx64, libhipblaslt and librocblas — the vendor libraries that other vendor libraries pull in (librocsparse needs librocblas and libroctx64; librocblas needs libhipblaslt). Order is load-bearing: HSA must precede HIP, or HIP's libhsa-runtime64.so.1 binds outside the bundle and the dlopen fails with undefined symbol: hsa_amd_queue_create, version ROCR_1.
• FOREIGN_LIBRARY_NAMES gains the vendor libraries, so a mixed process is reported instead of passing as clean.
• The out-of-process version probe is removed (_version_subprocess, _rocsparse_version_isolated, the cache global and its __init__ reset). With the libraries claimed correctly there is nothing left for it to protect against, and it was a lot of machinery for one cell of a table.
• versioninfo queries rocSPARSE in-process like every other library. The shared _ver helper now guards all of them, so a failing query shows err instead of taking down the whole report — previously none were guarded.

Result

On beverin (2x MI300A, artifact ROCm under uenv), before -> after:

	before	after
versioninfo rocSPARSE	err + misleading warning	5.0.0
librocblas / libroctx64 / libhipblaslt	uenv	bundle
foreign_libraries()	String[] (wrong)	String[] (correct)
preloaded libraries	41	45

Trade-off

A SIGSEGV inside rocsparse_create_handle (#920) again takes down the session rather than degrading to err. Two things make that acceptable: the preload fix removes the mixed-library cause that produced the crash here, and versioninfo already initialises HIP and enumerates devices in-process, so a genuinely broken install fails before the rocSPARSE row either way. try/catch still covers every non-fatal failure.

[simeonschaub]

I'm really not a big fan of introducing even more complexity just for the librocsparse version check. With #1028, we might want to revisit whether the check is truly needed anymore, since now the user would have to explicitly request AMDGPU.jl to use a broken system ROCm install for this to matter. I personally think it's not too unexpected that Julia would crash in such a situation, so I'd propose reverting #1001 altogether as part of #1028

[luraess]

Yeah, that's the other alternative, to remove all this machinery with the one caveat is that we may get inaccurate version for rocSPARSE if we ever encounter using it from system. I can do it here.

[luraess]

This is now mostly rolling back #1001 but adding some guards.