Align Split Miner gateway handling and SM2 lifecycle behavior with the reference implementation

[cRennert]

This PR revises the Split Miner implementation to better match the reference miner behavior across filtering, split discovery, join discovery, OR handling, lifecycle-aware SM2 logs, and BPMN reduction.

Key changes include:

• Replace the previous split/join handling with an Oracle-style split discovery and RPST/SESE-based join discovery.
• Add lifecycle-aware SM2 parsing with overlap-based concurrency detection for complex logs.
• Pin SM2 filtering to the reference eta = 1.0 behavior and make SM2 OR handling mandatory.
• Add gateway-map based inclusive-join handling, OR-split promotion, compact self-loop export, and optional BPMN gateway collapsing.
• Update Split Miner smoke/regression tests to reflect the new reference-aligned behavior.

[fit-alessandro-berti]

Dear @cRennert

Thanks for the contribution.

Can you double check on the following comments?

Possible Revisions Before Merge

• Restore or replace coverage for the removed SM2 improper-completion regression. The previous test asserted that A can repeat through a D-free loop in the paper-style example; this branch no longer satisfies that property. If the behavior change is intentional, document it explicitly in the PR and changelog.
• Consider compatibility shims for removed modules such as pm4py.algo.discovery.split_miner.heuristics, concurrency.refined, dfg_discovery.refined, and joins.classic, or call out the breaking import changes clearly.
• Add focused unit tests for gateway_map.py and rpst_tree.py, especially loop joins, fake entry/exit flows, token generator placement, and non-biconnected graphs.
• Add an explicit regression test showing that SM2 ignores user-provided eta and minimize_or_joins by design, so future reviewers do not mistake this for an accidental parameter bug.
• Include a short rationale or reference comparison for the large changes to classic filtering, split discovery, and join discovery, because the expected gateway counts changed from the prior tests.

[cRennert]

Possible revisions were addressed in d59dc67, 76eba37, and 40b3cf4 as follows:

Split Miner: faithful reimplementation of classic & SM 2.0
Classic Split Miner and Split Miner 2.0 are now ports of the reference Java tools (splitminer.jar and sm2.jar/MineWithSMTC), built to reproduce their output exactly rather than the idealized figures in the papers. Validated on the SM-Experiment corpus: classic is byte-identical (isomorphic, same gateway labels) to splitminer.jar on the deterministic real-life logs; SM 2.0 is byte-identical to sm2.jar on 9 of 10 reference logs (the 10th, 2015 BPIC, is non-deterministic in the Java tool itself).

Changed

• DFG filtering, concurrency oracle, Oracle split discovery, RPST/SESE join discovery, OR-join replacement and gateway collapse were reworked for byte-faithfulness. Because the target is the tool, not the papers, some hand-computed gateway counts in older tests changed — e.g. on the Augusto et al. (2019) running example the tool yields 2 AND-splits / 4 XOR-splits, not the 1/3 of Fig. 3c. Updated counts were each confirmed against the Java tool.
• SM 2.0 now pins eta = 1.0 and always runs its OR handling (replaceIORs = false + OR-split heuristic), matching the reference. The eta and minimize_or_joins arguments are accepted for API symmetry but ignored by variant='sm2' (covered by a dedicated regression test).

Removed (behavior change)

• The SM 2.0 improper-completion heuristic (which restructured an A→{B||C}→D→A loop so A could repeat through a D-free branch) was an approximation not present in the reference tool and has been dropped. Verified: Java SM 2.0 closes that loop after D, identical to this port. Its old regression test is replaced by tests/split_miner_2_loop_test.py, which pins the faithful behavior.
• Internal modules from the earlier approximate implementation were removed: …split_miner.heuristics, …concurrency.refined, …dfg_discovery.refined, …joins.classic. These were never part of the public API (pm4py.discover_bpmn_split_miner), but since they shipped in 2.7.22.4 each path now raises a descriptive ImportError pointing to the faithful replacement.

Added (tests)

• split_miner_2_loop_test.py — SM 2.0 loop handling (faithful, no D-free loop).
• split_miner_2_params_test.py — SM 2.0 ignores eta/minimize_or_joins by design.
• split_miner_internals_test.py — focused units for rpst_tree (polygon/bond/rigid/rejection) and gateway_map (OR-resolution, loop regions, token-generator placement, inclusive-retention).