FEAT: add qrules.workflow module

[grayson-helmholz]

• Closes Split up StateTransitionManager to facilitate use of intermediate results #305
• Closes Create struct and appropriate modification-functions for allowed_intermediate_particles #306
• Closes Struct and functions for allowed_interaction_types #307
• Closes Generate QNProblemSets without the StateTransitionManager  #284

✨ New features

• qrules.workflow exposes the pipeline as free functions: create_graph_settings(), create_problem_sets(), solve(), convert_to_particle_transitions(), and collect_reaction_info().
• create_qn_problem_sets() and find_solutions() cover the default use case in two calls, meeting at a QNProblemSetCollection that carries the final state, formalism, and particle selection from one to the other.
• AllowedIntermediateParticles, built with filter_intermediate_particles() and narrowed with select() and exclude(), which take the same name patterns and return a new selection.
• InteractionConfig, holding the interaction-type settings, the allowed types globally or per node, and the interaction determinators.
• SolutionContainer is public, because it is the boundary between convert_to_particle_transitions() and collect_reaction_info().

⚙️ Enhancements

• solve() takes strength-grouped QNProblemSets, so one filtered with filter_quantum_number_problem_set() can be fed straight back in.
• find_solutions() requires formalism and allowed_intermediate_particles for a plain dict of problem sets instead of guessing. Defaulting to "helicity" strips $L$ and $S$ from canonical-helicity solutions and mislabels ReactionInfo.formalism; falling back to the whole particle database matches intermediate states the problem sets were not built for. Both are still inferred from a QNProblemSetCollection.
• Multiprocessing pickles only the allowed-state property maps instead of the whole StateTransitionManager.

🔨 Maintenance

• The StateTransitionManager becomes a facade over these functions. Its public methods, arguments, and output are unchanged: tests/channels/ is untouched and a new test asserts that the pipeline reproduces the STM result for both formalisms.

📝 Documentation

• intermediate-results.ipynb is rewritten around the workflow functions. It keeps the dict_set_intersection() reduction from CSPSolver generates undesired edge_properties after filtering #301, without which the intermediate edges come back with the full PDG property maps.
• reaction.ipynb still teaches the StateTransitionManager exclusively; converting it is left as a follow-up.

Squash commit messages

* DOC: rewrite intermediate-results notebook around `workflow`
* ENH: require formalism and particles that cannot be inferred
* ENH: solve modified `QNProblemSet`s without the `CSPSolver`
* FEAT: add `create_qn_problem_sets` and `find_solutions`
* FEAT: add structs for intermediate particles and interactions
* MAINT: reduce `StateTransitionManager` to a workflow facade

[grayson-helmholz]