[WIP] Prototype original-parameter AdamW8bit state for FSDP1

[gss10282025]

Summary

Add an executable reference for keeping AdamW8bit state attached to original parameters under FSDP1. It gathers each original parameter's gradients, updates per-parameter shadow state, and copies the resulting values back into the flat shards.

Related to #2080.

This is a draft for discussing the state-layout contract. The reference replicates optimizer state and uses private FSDP1 metadata. A production version still needs a supported metadata interface and sharded state; checkpointing, AMP, and offload are outside this example.

Validation

Run with:

torchrun --standalone --nproc-per-node=2 examples/fsdp1_original_parameter_state.py --device cpu

Passed on PyTorch 2.13.0+cpu and bitsandbytes 0.50.2, with two CPU/Gloo ranks and two optimizer updates:

• Native AdamW8bit chooses FP32 state for the four small parameters with fine wrapping, and 8-bit state with coarse wrapping. The maximum absolute difference across the complete parameter-update vectors is 6.15e-5.
• The reference keeps those four states in FP32 under both layouts, and all parameter updates match exactly.
• The separate 8,192-element parameter keeps 8-bit state in both layouts.

Ruff lint/format checks and Python compilation pass. The example also accepts --device cuda; that command was not rerun for this PR.