docs(book): add DisTrO explainer chapter

[salch-cred]

Summary

Adds a new "Psyche in depth" chapter: DisTrO (explain/distro.md).

DisTrO was previously only referenced in passing — a box inside the decentralized training flow diagram in general-workflow.md and an optimizer stanza in the run-config docs — with no explanation of what it actually is or does. This chapter explains, from the Rust implementation in shared/modeling/src/distro.rs:

• The problem: gradient synchronization bandwidth makes conventional data-parallel training impossible between arbitrary internet nodes.
• The two ideas that fix it: transform-based compression (DCT + top-k per chunk) and residual accumulation, so small components dropped by top-k are deferred into later steps instead of being lost.
• A per-step walkthrough (mermaid diagram + numbered steps): decode peers' updates, apply aggregate, compute gradient, update the momentum residual (subtracting what was already sent to avoid double counting), encode, top-k, optional 1-bit sign quantization, broadcast over P2P.
• Where it fits: the optimizer handles all weight synchronization client-to-client; the Solana coordinator only handles the state machine, data assignments, and witness proofs — weight data never goes on-chain.
• Configuration: a table mapping compression_decay, compression_chunk, compression_topk, and quantize_1bit to their effect and their interaction with max_round_train_time.

Notes for reviewers

• All technical claims were verified against the code: Distro::generate / Distro::apply / error_correction in shared/modeling/src/distro.rs, the TransformDCT/CompressDCT encode/compress path, the Distro optimizer definition in shared/core/src/definitions.rs, and the quantization path in shared/modeling/src/trainer.rs (quantize_1bit → quantize_nozeros_tensor_to_boolean_sign).
• Only additions: one new page + one SUMMARY.md line.
• Happy to adjust tone, depth, or placement.