Fix: Enforce deterministic SDFG graph traversal and code generation

[kotsaloscv]

Motivation

When generating code via DaCe (e.g., as a backend for GT4Py/ICON4Py), clearing the Python cache between runs can result in structurally shifting generated code. While mathematically identical, these shifting instruction orders disrupt C++ compiler heuristics (such as register allocation and loop unrolling), leading to unpredictable performance regressions.

This non-determinism stems from graph traversal operations falling back on memory addresses or volatile generated UUIDs to break ties during dictionary and set iterations, which cascades into DaCe's internal adjacency lists and NetworkX topological sorts.

Solution

This PR introduces a deterministic sorting pass (sort_sdfg_alphabetically) that locks the internal memory layout of the SDFG before code generation by sorting elements based on their intrinsic semantic properties rather than their location in memory.

Specifically, this PR:

1. Semantic Node & Edge Keys: Introduces get_deterministic_node_key and get_deterministic_edge_key. To safely break ties between identical nodes, these functions evaluate graph topology (in/out degrees), interface semantics (connectors), loop parameters, and a stable MD5 hash of internal Tasklet code.
2. Strips Volatile Hashes: Implements a compiled regex (VOLATILE_STR_REGEX) to strip memory addresses, full UUIDs, and partial 8-character suffix hashes from node labels, connectors, and memlet data payloads.
3. Deep Graph Stabilization: Recursively sorts global symbol tables (_arrays, symbols), master node/edge dictionaries, and nested $O(1)$ adjacency lists (in_edges, out_edges) in-place.
4. Rebuilds NetworkX Backends: Tears down and deterministically rebuilds the underlying _nx multigraphs based on the newly stabilized registries.

Testing

Added tests/sdfg/deterministic_sort_test.py.
This test manually scrambles the internal dictionaries of an SDFG and SDFGState, applies the sorting pass, and asserts that the resulting adjacency lists and NetworkX nodes are forced back into a strict, predictable topological order based on the new semantic keys.

[tbennun]

This could have some performance implications. Could we make this configurable?

[philip-paul-mueller]

I have some comments ans suggestions.

[philip-paul-mueller]

There are still some things that needs to be checked, especially the global object.

[philip-paul-mueller]

There are a lot of small things.

[philip-paul-mueller]

I would change the CI configuration such that when auto optimize is activated (probably also when just simplification is enabled) that sorting is enabled.
You can do that by modifying .github/workflows/general-ci.yml.