Add tier-aware scheduling/weighting for heterogeneous training

[plugyawn]

Problem

The coordinator is tier-agnostic: data assignment is uniform across trainers and the centralized client ignores TrainingAssignment beyond logging. There is no enforcement of a target ratio (e.g., 50% prefix / 50% full) or tier-weighted contribution. Heterogeneous participation is therefore purely operator-managed rather than system-managed.

Refs:

• shared/coordinator/src/data_selection.rs (uniform batch assignment)
• architectures/centralized/server/src/app.rs (assignment = capabilities)
• architectures/centralized/client/src/app.rs (assignment stored, unused)

Expected

Tier-aware scheduling or weighting should be possible to hit a target mix of prefix/full training.

Possible Approach

• Add tier-aware weighting into assign_data_for_state (weighted by tier, token count, or configured ratio).
• Or, have coordinator emit assignments that override client tier for the round.
• Make clients actually apply TrainingAssignment to effective tier (or to loss weighting).

Acceptance Criteria

• Configurable target ratio for tier participation.
• Client uses assignment to influence effective tier or gradient weighting.
• Metrics/logging reflect actual tier mix per round.

[plugyawn]

Not sure what the right ideas forward are, but these are probably pertinent numbers: