[SYCL][Graph] Support sycl::queue::ext_oneapi_get_graph with native recording

[mmichel11]

• Queries UR to fetch queue associated graph handle from native API in order to support queue transitions from direct backend command submissions (e.g. L0)
• Introduces a map in context_impl linking UR graph handles to modifiable command graphs. queue::ext_oneapi_get_graph uses this map for lookup after calling urQueueGetGraphExp .
• Two test cases are added for command buffer and native recording paths. For native recording, minimum driver versions are added to ensure support for underlying L0 API and L0 bugfixes for fork-join.

[adamfidel]

Can we split the UR_RESULT_SUCCESS branch from the !UrGraphHandle branch? If the UR call fails (e.g. unsupported feature) the user will see an error saying saying "can only be called on recording queues," which might be a bit confusing.

[mmichel11]

Good point. I switched the checks. The UR spec returns UR_RESULT_ERROR_INVALID_OPERATION if the queue is not recording (and sets the pointer to null pointer). I switched to check for this first to throw "can only be called on recording queues". If there is any other error code, then it is internal and should throw with errc::runtime instead of invalid.

[mmichel11]

It ends up being a bit more complicated to return the correct error codes due to backends that don't support graphs and driver versions where native get graph is unsupported but other graphs operations are.

I fixed this by adding a isNativeRecording() query. It lets us know which exception should be returned for drivers where get graph is not supported.

[adamfidel]

LGTM!

[mmichel11]

@sergey-semenov @intel/llvm-reviewers-runtime Can you take a look?

[mmichel11]

@intel/llvm-reviewers-runtime Friendly ping for review. There are some workloads I would like to get this in for.