Borrow bf16 [M,K] and axis-0 1MAD weights; return owned lazy [M,N].
M=5..8, T=256/L=12, k=2/3/4. Match MLX's non-NAX split-K matrix regime;
other shapes and hardware require their own dispatch and full-model gates.
Peak workspace is SPLITSMN float elements, with no dense weight buffer.
Borrow bf16 [M,K] and axis-0 1MAD weights; return owned lazy [M,N]. M=5..8, T=256/L=12, k=2/3/4. Match MLX's non-NAX split-K matrix regime; other shapes and hardware require their own dispatch and full-model gates. Peak workspace is SPLITSMN float elements, with no dense weight buffer.