Fix GaLore 8-bit updates with ZeRO CPU offload

[sdjasj]

What changed

• stage each CPU-offloaded distributed GaLore parameter shard and gradient on the accelerator immediately before its bitsandbytes update
• move non-paged 8-bit optimizer state and GaLore projector state with the parameter
• return the updated parameter, gradient, and optimizer state to CPU after every parameter update
• add a two-step distributed regression that verifies the offloaded state remains on CPU between steps

Why

bitsandbytes only provides GPU update kernels. LowLevelZeroPlugin(cpu_offload=True) placed the master parameter, gradient, quantized states, maps, and block maxima on CPU, so the kernel rejected every input as non-GPU.

Staging one parameter at a time preserves CPU offload between updates and avoids materializing all offloaded optimizer state on the GPU at once.

Impact

GaLoreAdamW8bit can now train with FP32 LowLevelZeroPlugin CPU offload.

Fixes #6399.

Validation

• python -m pytest tests/test_optimizer/test_dist_galore.py -q (4 GPUs; 1 passed)
• the regression covers two consecutive CPU-offload optimizer steps and verifies master parameters and non-paged optimizer state return to CPU
• compileall and git diff --check: passed