[CI][E2E] Run E2E tests on NVIDIA with prebuilt binaries in precommit

[ayylol]

Reduces time for precommit nvidia job from around 20m to around 5m

[sarnex]

Current NVIDIA job takes ~20m and with this PR it takes ~5. Great speedup!

[aelovikov-intel]

Before (https://github.com/intel/llvm/actions/runs/13262954234/job/37025407654):

Testing Time: 1218.09s

Total Discovered Tests: 2274
  Unsupported      : 1191 (52.37%)
  Passed           : 1061 (46.66%)
  Expectedly Failed:   22 (0.97%)

After (https://github.com/intel/llvm/actions/runs/13266214639/job/37037510106?pr=16813):

Testing Time: 183.58s

Total Discovered Tests: 2274
  Unsupported      : 1199 (52.73%)
  Passed           : 1053 (46.31%)
  Expectedly Failed:   22 (0.97%)

What is the reason for 8 new UNSUPPORTEDs ? Where will we be testing those?

[ayylol]

What is the reason for 8 new UNSUPPORTEDs ? Where will we be testing those?

Those extra unsupported are the following tests:

DeviceLib/math_override_test.cpp
Reduction/reduction_dynamic_span.cpp
Regression/2020-spec-constants-debug-info.cpp
Regression/optimization_level_debug_info_intopt.cpp
Regression/optimization_level_debug_info_specopt.cpp
bindless_images/examples/example_6_import_memory_and_semaphores.cpp
Basic/interop/interop_all_backends.cpp
Adapters/adapter-release.cpp

They are all marked as unsupported on the run-only side because they only have build lines. They are being ran on the build step only. The only one of these where this is incorrect behaviour is the Adapters test, since it has a single run line that executes sycl-ls, I can change that in a separate pr.

[YuriPlyakhin]

Joint Matrix change LGTM

[sarnex]

Seems approved by everyone, @ayylol ready for merge?