Simplify normalize_regime_transition_probs and other cleanups

[hmgaudecker]

Address some parts of #187. This simplifies the code a bit. I did not benchmark, but I do not see where it could lead to a slowdown.

[timmens]

TBH, I am confused by the PR description / title and the code changes.

1. We now have two normalize regime transition probs functions. The one that is tested is not used in the code base, and the one that is used is not tested.
2. Given your statement that the regime passing is simplified, I would've expected to see changes in the models. This is actually done in #213 right?

[timmens]

Issue #187 asks to:

Use the Array-return interface also internally [for regime transitions]

The issue notes that after #186, the user-facing interface supports returning Arrays directly, but internally the code still relies on dict[str, float | Array] format.

What the PR actually does:

1. Changes type annotations from dict to MappingProxyType and list to tuple
2. Refactors normalize_regime_transition_probs to work with MappingProxyType[str, Array] instead of dict[str, float | Array]
3. Various cleanup and refactoring

What's still dict-based internally:

1. In src/lcm/simulation/util.py:160-168:
   regime_transition_probs: MappingProxyType[str, Array] = (
   internal_regime.internal_functions.regime_transition_probs.simulate(...)
   )
2. In src/lcm/Q_and_F.py:138-142:
   regime_transition_prob: MappingProxyType[str, Array] = (
   regime_transition_prob_func(...)
   )
3. normalize_regime_transition_probs still takes and returns MappingProxyType[str, Array]
4. draw_key_from_dict still consumes a dict of arrays

The core internal representation remains dict[RegimeName, Array] (now wrapped in MappingProxyType). To actually fix issue 187, the internal
interface would need to change to use a single Array with shape (n_points, n_regimes) instead of a dict mapping regime names to probability
arrays.

---

While this is something we could do, I actually don't think it has a high priority right now.