Two small example bugs: wgrad split-k dumps the wrong tensor as Computed; FMHA diff-checker launches with block/grid swapped

Description

Two small example defects spotted while running the examples on an RTX 5060 Ti (sm_120a, CUDA 12.8.1):

1. wgrad split-k workspace dump labels the wrong tensor as "Computed"

examples/30_wgrad_split_k/30_wgrad_split_k.cu ~line 621:

output_workspace << "Computed = \n" << tensor_c.host_view() << std::endl;

tensor_c is the zero-filled bias/filter-C input; the kernel output is tensor_d. With --save-workspace, anyone debugging gets a zeros dump labeled as the computed result.

Fix: print tensor_d.host_view().

2. Block/grid swapped in the FMHA reference diff-checker launches

examples/77_blackwell_fmha/reference/reference_abs_error.hpp (~lines 161-163 and 257-259) and examples/88_hopper_fmha/reference/reference_abs_error.hpp (~line 106):

dim3 block(256, 1, 1);
dim3 grid(1024, 1, 1);
reference_abs_diff_kernel<<<block, grid>>>(...);

This launches 256 blocks x 1024 threads instead of the intended 1024 x 256. Harmless for correctness here (grid-stride loops plus blockDim-sized shared reductions), but it teaches a wrong launch idiom in two widely copied files.

Fix: <<<grid, block>>>.

Inspect the indicated lines in examples/30_wgrad_split_k/30_wgrad_split_k.cu and the reference_abs_error.hpp files under examples/77_blackwell_fmha and examples/88_hopper_fmha. Compare the named tensors and kernel launch arguments with the surrounding code, then run the affected examples, including the split-k workspace dump, to verify the reported output and launch behavior are corrected.