[Orchestration/Performance] Transient memory scoping & transformation

[FlorianDeconinck]

With the introduction of transient in the orchestration pipeline we have the basis for two central cache-related optimization:

• Refining dimensionality of temporaries (RefinedNestedAccess pass in DaCe)
• Remove/replace fields that are no longer useful after reduction and merge (DeadFlowElimination in DaCe)

One issue is that all fields are considered lifetime.Persistent in the DaCe -> GT4Py bridge. Technically, those should be tagged as lifetime.Scope when orchestrating. lifetime.Persistent allows us to properly pool memory, a feature we need to keep. So we will need to re-scope to lifetime.Scope at first, then once optmization is done, de-scope back to Persistent and pool.

Flow of optimization for transient:

1. Re-scope transient in to lifetime.Scope (gt4py-dace bridge, potentially only for orchestration)
2. Apply merging of IJ or K loops
3. Apply RefineNestedAccess to reduce dimensionality
4. Simplify (scalarize & DeadFlowElimination)
5. For any transient still alive, de-scope to lifetime.Persistent and flag has pool

[FlorianDeconinck]

Step 1 and 2 have been tested in a small example.

[romanc]

DaCe 2.0 will have a pass to detect and transform what gt4py knows as "K-Caches".