Add per-step reward field to Action and Observation schemas

[henryjcee]

Summary

• Adds an optional reward: float | None field to the base Action class (schema/action/action.py), inherited by ApiAction, CodeAction, and MessageAction
• Adds an optional reward: float | None field to the base Observation class (schema/observation/observation.py), inherited by TextObservation, WebObservation, and ImageObservation
• Updates schema/SCHEMA.md to document the new field on both base classes

Motivation

ADP currently has no mechanism to attach reward signals to individual trajectory steps. This makes it difficult to use ADP-formatted data for reinforcement learning, where per-step rewards are a core primitive.

This change adds reward as a first-class optional field on every action and observation, allowing datasets to record the reward received at each step of a trajectory. It's conceivable that some RL settings may provide reward with an observation or at action time, this change supports both approaches.

Design notes

• reward defaults to None — fully backwards-compatible, all existing sample_std.json files validate without modification
• Modelled as a plain float scalar (not a distribution or vector) to keep the schema simple and composable

Tests

• pytest tests/test_standardized_schemas.py — all 33 datasets pass
• Full test suite — 136 passed, 0 failures

I don't think new tests are required by this but happy to add if useful.