tune(dsv4): add 16384x1536 & 7168x768 a8w8 blockscale bpreshuffle, prefer ck over asm+splitK where within 1us (gfx950) (#4326)

Tuned on gfx950 (MI355X, cu_num=256):

1. Add missing tuned configs for two shapes that fell back to default:
   - N=16384 K=1536 and N=7168 K=768, full standard M sweep (1..65536).
   - Small M (1/4/8/16) of 16384x1536 keep the existing faster triton entries.

2. For shapes that resolved to an asm kernel with splitK enabled, re-tune with
   ck/cktile only and switch to ck where it is faster or within 1us of asm
   (24 shapes). Biggest win: M=800 N=7168 K=768 asm 27.6us -> ck 14.0us (1.97x).
   Shapes where asm+splitK stays clearly faster (skinny-N / large-K such as
   768x7168, 7168x16384, 6144x7168) are left on asm.

Author: valarLip