fix[next-dace]: skip partially written temporaries in DistributedBufferRelocator

[edopao]

Problem

The stree lowering of concat_where emits a transient array that is partially written in two consecutive SDFG states, and the second state also copies the transient to a global array (its "write-back" state). DistributedBufferRelocator._find_candidates scans such write-back states and assumes the transient's write-back node is a pure sink, requiring write_back_state.in_degree(write_back_node) == 0. On the concat_where pattern the write-back node also has an incoming edge (the partial write in the same state), so gt_simplify crashed with an AssertionError (observed e.g. on ICON OIR programs). Per ADR-018 a transient must be written within a single state, but the relocator must not crash while other passes handle such patterns.

Fix

A temporary whose write-back node is also written in the write-back state (i.e. state.in_degree(temp_storage) != 0) is simply not treated as a relocation candidate and is left untouched.

Tests

• New regression test test_distributed_buffer_partial_write_back builds an SDFG with the partial-write-in-write-back-state pattern and asserts that gt_reduce_distributed_buffering no longer crashes, records no DistributedBufferRelocator relocation, and leaves the state intact.
• The full transformation_tests/test_distributed_buffer_relocator.py suite passes (8 tests).