feat(optimizer): add opt-in packed 4-bit AdamW states

[xsuler]

Summary

• add opt-in --adam-4bit / adam_4bit optimizer selection for the CUDA backend
• store both Adam moments as packed 4-bit nibbles with parameter-local B128 scales (about 1.06 bytes/parameter including FP32 scale metadata for large tensors)
• use the signed dynamic-exponent first-moment map and zero-excluding second-moment map from Memory Efficient Optimizers with 4-bit States
• add fused CUDA updates that operate directly on packed state, plus bounded block-wise CPU reference updates and versioned checkpoint loading
• make the existing 8-bit optimizer parameter-local/block-wise so outliers cannot set one bucket-global scale and amplify the effective step at the same learning rate
• skip only a block containing non-finite values instead of poisoning neighboring packed state
• expose the option through typed trainer/engine configs, CLI summaries, and dashboard command construction

Validation

• ruff check and ruff format
• Python compileall
• git diff --check
• CPU tests cover packing/storage budget, quantizer maps, checkpoint round trips, disk offload, FP32 tracking, non-finite block isolation, configuration/CLI routing, and the 8-bit same-learning-rate regression
• CUDA parity tests are included for CI; this development host has no CUDA toolchain/device, so the fused extension was not built locally

Notes

The model checkpoint format is unchanged. Only the optimizer checkpoint carries the explicit 4-bit format version and block size.

Closes #535