Fix stale device global initializers on RTC program reuse

[kweronsx]

Prevent stale MDeviceGlobalInitializers entries from colliding with new entries when program handles are reused in RuntimeCompiled kernel loops. Implement explicit cleanup via removeDeviceGlobalInitializer() during program teardown.

[sergey-semenov]

Please extend sycl/unittests/program_manager/Cleanup.cpp (or another unit test, if more appropriate) to cover this.

[kweronsx]

This change is required to fix this test failure: https://github.com/intel/llvm/actions/runs/24240135791/job/70775597393?pr=21728

Please extend sycl/unittests/program_manager/Cleanup.cpp (or another unit test, if more appropriate) to cover this.

I have extended the sycl_device_globals test to catch this: (#21728)

Additional info:
When a program goes out of scope, its entry in the device global initializer map is now removed. Previously, if the adapter reused the same program handle for a newly created program, the stale entry (already marked as fully initialized) would be found by initializeDeviceGlobals and initialization would be silently skipped. This means device globals would retain stale state from the previous program lifetime. Now the entry is cleaned up on program teardown, so the next time the program is built and run — even inside a loop — its device globals are properly re-initialized.

[sergey-semenov]

I have extended the sycl_device_globals test to catch this

That's an E2E test, but since we have a dedicated unit test for the cleanup mechanism when removing images, I think it should be updated to include this too.