Validation should reject direct tasklet/code-node -> tasklet|NestedSDFG edges without an intermediate access node

## Summary

DaCe currently allows a direct edge between two tasklets (or from a tasklet/code node/library node into a NestedSDFG) that carries named data, with no AccessNode in between. This makes dataflow implicit and causes real bugs: array/scalar elimination inside simplify can remove a value that only "exists" on such an edge, and `InlineSDFG` then crashes trying to propagate it upward, since it expects a producing AccessNode.

## Where this came up

Discussed on PR #2586: https://github.com/spcl/dace/pull/2586

- `_can_drop_shared_paths` checks whether a shared nested-SDFG connector "without a source (or sink) access node" can be dropped. I think the in/out connector on a NestedSDFG should always trace back to an AccessNode through the memlet path (e.g. AccessNode -> MapEntry -> NestedSDFG), so a missing source/sink access node looked confusing.
- Separately, a reported failure in `indirection_test.py` (`test()`) showed the same root cause: with a scalar (`__bla_value`) sitting directly on a tasklet -> tasklet edge (no AccessNode), array/scalar elimination removes it from the nested SDFG, and `InlineSDFG` then hits a hard crash (KeyError) trying to propagate it back up, because it assumes an AccessNode is present. The same issue would apply to arrays, not just scalars.

## Proposal

Treat a direct code-node -> code-node edge (tasklet -> tasklet, tasklet -> NestedSDFG, code node/library node -> NestedSDFG, etc.) carrying named data as invalid IR: there should always be an intermediate AccessNode making the dataflow explicit, i.e. code node -> AccessNode -> code node|NestedSDFG. This should be enforced in SDFG validation rather than silently tolerated, so passes (like the array/scalar elimination that is part of simplify) can rely on the invariant instead of being that ones causing crashes downstream.

This is a design change with tradeoffs (there may be existing graphs/tests relying on tasklet -> tasklet edges), so it likely needs discussion before implementation.

Review the SDFG validation logic and the failure described in indirection_test.py's test(), then trace how simplify's array/scalar elimination reaches InlineSDFG. Check whether existing graphs or tests rely on direct named-data edges and resolve the proposed invariant across tasklet, code-node, and NestedSDFG cases. Done means the validation behavior and affected failure paths are covered by tests.

Written by the indexing model from the issue text.