DeepSpeed

d4d7cadd - fix: materialize strided TP q/k/v views for the in-place GDN recurrent step

fix: materialize strided TP q/k/v views for the in-place GDN recurrent step TP block-sliced value arrived as a strided view of the fused in_proj buffer, silently failing the contiguity gate and routing every decode step to FLA's out-of-place kernel (24 layers x state copy-back per step = the 27 memcpy/st and half the recurrent state traffic seen in the A100 profile). Materialize the tiny views instead. Signed-off-by: Guokai Ma <guokai.ma@intel.com>