RL Float32 Numeric Checks (Umbrella)

This is the stable import for TorchLean's explicit binary32 RL diagnostics. The implementation is split by concern:

• Types: IEEE32Exec/Interval32 aliases, boundary casts, and checked scalar primitives;
• Returns: checked discounted backups and fixed-horizon returns;
• Advantage: checked TD residuals, GAE(λ), and advantage normalization;
• PPO: checked importance ratios and clipped PPO objective pieces;
• Intervals: outward-rounded interval enclosures for return/GAE/PPO diagnostics.

The public namespace remains Runtime.RL.Numerics.Float32, so existing downstream imports and names continue to work while the source layout is easier to review.

References:

• IEEE 754-2019 (binary32 arithmetic): https://doi.org/10.1109/IEEESTD.2019.8766229
• IEEE 1788-2015 (interval arithmetic API/semantics): https://doi.org/10.1109/IEEESTD.2015.7140721
• Goldberg, “What Every Computer Scientist Should Know About Floating-Point Arithmetic” (1991): https://doi.org/10.1145/103162.103163
• Sutton and Barto, Reinforcement Learning: An Introduction (2nd ed.).
• Schulman et al., "High-Dimensional Continuous Control Using Generalized Advantage Estimation" (2015), and "Proximal Policy Optimization Algorithms" (2017).