Add grid-stride + ROCm cap to grad_mean kernel

[q10]

Summary:
The MEAN-pooling TBE backward launches grad_mean{,_vbe}_kernel with
grid = div_round_up(total_B, kMaxThreads / grad_mean_warp_size) and
block = dim3(grad_mean_warp_size, kMaxThreads / grad_mean_warp_size)
(total_B = B * T), so total threads ~= total_B * grad_mean_warp_size exceeds
the HIP 2^32 threads-per-launch limit on ROCm for large total_B.

Cap the launch with utils::cuda::cap_grid_dim_x(..., OverflowOnly) and add a
ROCm grid-stride loop over b_t to the kernel (no internal early-returns; single
guard converted to the loop bound). No-op on CUDA.

Reviewed By: cthi

Differential Revision: D113351689

[meta-codesync]

@q10 has exported this pull request. If you are a Meta employee, you can view the originating Diff in D113351689.