[OpenMP][libomptarget] Enable lazy device initialization

[doru1004]

Enable lazy device initialization. This addresses this issue: #75677

This is a runtime feature which can be enabled using the following environment variable when running an OpenMP program:

LIBOMPTARGET_LAZY_DEVICE_INIT=1

[jdoerfert]

I mark this as requests changes until we resolve the conversation on the issue. It looks like we are still seeing different profile results, right?

[doru1004]

Whether we can reproduce upstream or not I think we should allow for the user to choose which way the device initialization happens eager or lazy.
As a side effect it will provide a resolution to the issue hence the link to the issue.

[jdoerfert]

Whether we can reproduce upstream or not I think we should allow for the user to choose which way the device initialization happens eager or lazy. As a side effect it will provide a resolution to the issue hence the link to the issue.

Could you clarify why we would carry more code to give the user this choice if upstream does not experience the same issue?

[doru1004]

Whether we can reproduce upstream or not I think we should allow for the user to choose which way the device initialization happens eager or lazy. As a side effect it will provide a resolution to the issue hence the link to the issue.

Could you clarify why we would carry more code to give the user this choice if upstream does not experience the same issue?

It's not about the amount of code (which is minimal) it's about providing a choice. There's no clear choice between the two strategies so we need to provide both. Moreover we would find this addition useful so even more reason to add it.

[doru1004]

Can confirm that this patch does indeed help with PCIe interconnects.