Introduce a `DiscreteMechanism` base class in base.py that implements the
Select-Measure-Estimate paradigm from McKenna et al. (2021). Each step of
the pipeline is a separate overridable method, enabling code reuse across
mechanisms that differ primarily in the *select* step.

New abstractions:
- `DiscreteMechanism`: Base dataclass with `configure()` for budget setup.
  `__call__` owns the fixed pre/post scaffolding (measure one-way, compress,
  generate, package result); `_run` is the overridable middle where the
  estimation strategy lives. One-shot mechanisms override only `_select`;
  adaptive mechanisms override `_run`.
- `GaussianMarginalMeasurement`: A `DPMechanism` that encapsulates
  Gaussian noise addition to marginal queries with proper `dp_event`.
- `_one_way_cliques` hook: Subclasses can restrict which one-way
  marginals are measured (AIM/AIM-GDP use workload-specific columns).

Mechanism migrations:
- Independent: one_way_budget_fraction=1.0, _select -> [], delegates dp_event
- Direct: one_way_budget_fraction=0.0, _select -> prespecified queries
- MST: _select delegates to existing exponential mechanism helper
- AIM: Overrides _run for the adaptive annealed loop, reuses building blocks
- AIM-GDP: Same _run pattern as AIM but in GDP budget space
- SWIFT: Overrides _select, _precompile, _measure, _estimate, _generate
  for per-clique budget allocation and junction tree precompilation

PiperOrigin-RevId: 944103007