Add comprehensive Parallelisms white paper (multi-engine, topology-aware draft)

[dzmitry-huba]

Motivation

• Provide a comprehensive, topology-aware reference that documents parallelism strategies, notation, cost models, and engineering recipes for training and serving large models across device meshes.
• Capture practical decision rubrics, failure modes, diagnostics, and performance primitives for a wide range of axes including DP, TP, PP, SP/CP, MoE, hybrid meshes, inference, and CUDA/platform techniques.
• Structure the material as independent engine blocks so each subsection is self-contained for specialist review and incremental adoption.

Description

• Add new documentation file papers/parallelisms_whitepaper.md containing an expanded, multi-engine draft (~3300 lines) with global notation, common equations, and per-subsection Engine blocks covering sections 1–13 and many detailed primitives and recipes.
• Document primitives and cost models for collectives (AR, AG, RS, A2A), memory/communication accounting, layout/reshard mechanics, and topology-aware scheduling guidance.
• Include end-to-end execution recipes, case studies, bibliography and verification notes, and explicit decision rubrics to guide selection and composition of parallel axes.
• Organize the content to support modular review, reuse, and future extraction of actionable implementation items.

Testing

• No automated tests were added for this documentation-only change.
• The new markdown file was committed to the repository and will be validated by existing documentation CI (linting/build) where configured.

---

Codex Task