Fix build with current DPC++

[fknorr]

Currently the continuous DPC++ build is broken on three issues:

• oneAPI is incompatible with mimalloc (see Use of dlopen(RTLD_DEEPBIND) breaks custom memory allocators oneapi-src/unified-runtime#803)
• malloc_shared produces host-inaccessible alloctions on Intel Arc GPUs
• Intel Arc GPUs do not support device-to-device copies.

This PR works around all of these issues in a minimally invasive way.

[github-actions]

Check-perf-impact results: (9e900a306dc9f17e4a27439205a7680c)

❓ No new benchmark data submitted. ❓
Please re-run the microbenchmarks and include the results if your commit could potentially affect performance.

[PeterTh]

Nice!
Well, the amount of stuff which simply doesn't work on Intel right now isn't so nice, but it's very nice to be back in green CI land.