b6691

ggml webgpu: actually add softmax, fix rms_norm offset (#16400)

* implement soft_max

* Fix soft_max data race

* Temporary fix, wait on each submit