fix(sampling): sampling span missing for strategies that override sample() #1487

Background

The sampling span added under #1050 is emitted by SamplingTracingPlugin subscribing to the sampling_loop_start / sampling_iteration / sampling_repair / sampling_loop_end hooks. Those hooks are dispatched inline inside BaseSamplingStrategy.sample() (mellea/stdlib/sampling/base.py).

Any strategy that overrides sample() rather than reusing the base loop therefore emits no sampling span, no iteration/repair events, and no error-path closure:

• BudgetForcingSamplingStrategy.sample() (budget_forcing.py:101) — own loop, no hook dispatch.
• SOFAISamplingStrategy.sample() (sofai.py:565) — own S1/S2 loop, no hook dispatch.
• BaseMBRDSampling.sample() / majority-voting (majority_voting.py:102) — fans out to super().sample() per inner sample, so it emits one sampling span per inner sample with no enclosing span for the majority-vote operation. The nesting is incidental, not designed.

Why this matters

sample() is a public, override-designed method, but sampling telemetry is coupled to one specific method body rather than to the strategy boundary. A strategy silently drops out of tracing with no signal, and the majority-voting case shows the coupling already produces surprising span shapes. This is the same class of gap PR #1264 / #1183 fixed for generate_from_raw, where inline hook firing duplicated across backends was moved into a base-class wrapper (generate_from_raw fires hooks, _generate_from_raw_impl is the override).

Scope

Mirror the generate_from_raw fix:

1. SamplingStrategy.sample() becomes a concrete base method that mints the sampling_id and fires the sampling lifecycle hooks (start/iteration/repair/end) with error-path closure.
2. Add an abstract _sample_impl (or equivalent) that strategies override for their actual loop.
3. Move BaseSamplingStrategy's current loop body into its _sample_impl; migrate BudgetForcing, SOFAI, and majority-voting to the same pattern so each emits exactly one enclosing sampling span.
4. Add firing-site tests using a mock strategy that overrides only _sample_impl.

Acceptance criteria

• Every SamplingStrategy (base loop, budget-forcing, SOFAI, majority-voting) emits exactly one enclosing sampling span per top-level sample() call.
• Iteration/repair events and error-path closure fire regardless of which strategy is used.
• No inline sampling_loop_* hook dispatch remains in individual strategy sample() bodies.
• Firing-site tests cover the hooks against a mock strategy.