[SYCL][Coverity] Fix MISSING_LOCK coverity hits in the program manager

[ianayl]

This PR guards accesses to m_DeviceImage and m_VFSet2BinImage caches/maps in the Program Manager using m_KernelIDsMutex in situations where multiple threads may be accessing the maps at the same time.

I am opting to use m_KernelIDsMutex here as both maps are commonly accessed while holding said mutex: In fact, a lot of maps/caches in the program manager use m_KernelIDsMutex (This is not the most optimal solution, but that is out of the scope of this PR, see #18165)

This PR addresses Coverity hits that needs to have fixes submitted before the code cutoff.

[ianayl]

@aelovikov-intel Sergey's on vacation and wont be able to review in time, is there someone that can review it sooner?

[aelovikov-intel]

@aelovikov-intel Sergey's on vacation and wont be able to review in time, is there someone that can review it sooner?

Maybe @KseniyaTikhomirova or @steffenlarsen ... I definitely don't think that going with a hack is the right thing to do here. Also, I don't know if there might be some other locks up the call stack that might make this a false positive.

[ianayl]

Would you prefer an additional mutex? I thought about adding another mutex, but decided against it since every other access to these maps has used m_KernelIDsMutex anyway. Additionally, a lot of maps and caches in the program manager use m_KernelIDsMutex even if they are not specifically caching/mapping kernel IDs, so I figured it wasn't that much of a hack.

I checked the call stack up to getOrCreateKernel and didn't see another lock, although it is not impossible that I may be wrong.

[aelovikov-intel]

#16836 seems to be related.

[aelovikov-intel]

IMO, you need to reword your description then. Be more assertive that this is the right change that follows what pre-existing codes does and drop any hesitation from it.

[steffenlarsen]

I think this is fine, but please open a tracker to improve the locking. Also, I agree with @aelovikov-intel that the locking in removeImages should be done outside the top-level loop, with a check that there are binaries before locking (or an early exit.)

[ianayl]

Tracker created in #18165

[steffenlarsen]

LGTM!

[ianayl]

@intel/llvm-gatekeepers PR ready for merge, thanks!