[SYCL][Graph] Reuse recording queue when available for finalize

[reble]

There are two issues with the current implementation:

• Creating a placeholder queue is adding overhead at finalize() to common path.
• Queue properties are set to default when no queue can be reused at Graph finalize().

[aelovikov-intel]

I think this should return std::shared_ptr instead (or maybe a reference to weak). Creating a copy of the weak_ptr requies atomics and isn't much faster than creating a std::shared_ptr (if at all?, AFAIK). As such, creating a copy here followed by the lock/dtor at the use side is like triple atomics vs a single one when returning a shared pointer.

I might be totally wrong here, but that's my best understanding of how those things work.

[reble]

Good catch and thanks for your feedback!
Let's use shared_ptr since it works better with the common case.
I don't think reference to weak is easily applicable, because we don't necessarily have a weak_ptr to return.
Even though I like the idea.

[mmichel11]

LGTM

[github-actions]

@intel/llvm-gatekeepers please consider merging