ggml-et: Retrieve kernel errors

[vidas]

Check kernel errors (return values).

Fail and complain loudly if kernel execution fails. Skips empty contexts (you don't want ~1000 debug blocks if one thread failed)(do you?).

@glguida I just remembered that I have this commit in my "working branch" - only checks for user errors, but could be a starting point to understand other kernel launching issues. Didn't find stream callbacks necessary in this case.

[glguida]

So this IIUC -- away from code -- implies error on execution of kernel.

For runtime errors when launching the kernel by the MM, we need this:

https://github.com/aifoundry-org/et-platform/blob/022221990ca713f946eaa4379817cb59bb852be4/esperanto-tools-libs/src/Runtime.cpp#L170

At the moment, these kind of runtime errors are ignored. When the callback is set, we can generate an exception, which will be catched by the try block.

In general this patch looks fine, but we need to arm the runtime callback to generate exceptions too.

[glguida]

LGTM, now tested. This doesn't stop execution though, so it seems like we need the callback.

[glguida]

It would be very useful to print the device number here, too.

While this code is correct, I wonder if the Stream Error callback with exception is actually the right mechanism, instead of manually parsing the error.