[AutoEP] Run staged expert weight gradients with ZeRO-3
ZeRO-3's gradient hook now skips a parameter the backward delivered no
gradient for, and the grouped-GEMM expert paths copy a weight that ZeRO-3
gathered to an address torch._grouped_mm does not accept. The deferred
expert function saves detached aliases of the weights, since ZeRO-3 swaps a
parameter's data after the forward.
Correct accumulation across windows additionally needs a parameter's first
gradient of a window to overwrite its ZeRO-3 gradient partition, which
#8981 provides.
Co-Authored-By: Claude Fable 5.2 <noreply@anthropic.com>