feat(optimizer): complete dynamic AdamW8bit state routing

[xsuler]

Summary

• implement paper-compatible signed/unsigned dynamic 8-bit codebooks in CUDA and MLX
• keep token-embedding Adam moments in FP32 using explicit role/identity routing, with tied-parameter deduplication and explicit precision overrides
• support mixed-state checkpointing, DP sharding, CPU/disk offload, clearing, and state-memory metrics without changing pretrained model forward behavior
• preserve AdamW4bit behavior while sharing the optimizer base

Validation

• .venv-test/bin/python -m pytest -q tests/test_adamw_8bit_blockwise_cpu.py tests/test_adamw_4bit_cpu.py — 24 passed
• targeted MLX AdamW8bit tests — 5 passed
• targeted Ruff checks/formatting, py_compile, and git diff --check

CUDA compilation/runtime and the issue three-seed training/throughput experiments require the project GPU environment and were not run locally.

Per maintainer direction, this intentionally does not support optimizer checkpoints produced by the previous linear AdamW8bit implementation.

Addresses #541