[BUG] `Aborted (core dumped)` on incorrect copy layouts

Hi!

I came across a minor bug while using the new CuTe DSL. When incorrectly specifying a tv layout, CuTe DSL segfaults instead of giving an error.

Reproducer:

import torch
import cutlass
import cutlass.cute as cute
from cutlass.cute.runtime import from_dlpack

M = 64
device = torch.device("cuda")
A = torch.randn(M, M, dtype=torch.float32, device=device)
A_tiler = (M, M)

@cute.kernel
def test_kernel(
    mAT: cute.Tensor,
    sA_layout: cute.Layout,
    copy_A: cute.TiledCopy
):
    tidx, _, _ = cute.arch.thread_idx()

    gAT = cute.local_tile(mAT, tiler=A_tiler, coord=(1, 1))

    smem = cutlass.utils.SmemAllocator()
    sA = smem.allocate_tensor(mAT.element_type, sA_layout, byte_alignment=16)

    thr_copy_A = copy_A.get_slice(tidx)

    tAgAT = thr_copy_A.partition_S(gAT)
    tAsA = thr_copy_A.partition_D(sA)

    cute.copy(copy_A, tAgAT, tAsA)
    return


@cute.jit
def test_fn(mA: cute.Tensor):

    mAT_layout = cute.make_layout((M, M), stride=(1, M))
    mAT = cute.make_tensor(mA.iterator, layout=mAT_layout)

    sA_layout = cute.make_layout((M, M), stride=(M, 1))
    etype_mA = mA.element_type
    smem_size = cute.size_in_bytes(etype_mA, sA_layout)

    tA = cute.make_layout((16, 16), stride=(16, 1))
    vA = cute.make_layout((1, 1))

    atom_async_copy_A = cute.make_copy_atom(
        cute.nvgpu.cpasync.CopyG2SOp(),
        etype_mA, num_bits_per_copy=etype_mA.width,
    )

    # should be cute.make_tiled_copy_tv, "Aborted (core dumped)" as is
    tiled_copy_A = cute.make_tiled_copy(atom_async_copy_A, tA, vA)

    test_kernel(mAT, sA_layout, tiled_copy_A).launch(
        grid=(1, 1, 1),
        block=(256, 1, 1),
        smem=smem_size
    )

A_tensor = from_dlpack(A)

gemm = cute.compile(test_fn, A_tensor)
gemm(A)
torch.cuda.synchronize()

Here, if I use make_tiled_copy_tv (as I actually intended) it works. However, if I use make_tiled_copy it core dumps. I think that it would be nice if using make_tiled_copy, while incorrect, would produce an error message instead of an Aborted (core dumped). It's a bit tricky to debug core dumps 😅

Thanks!

Start with the make_tiled_copy and make_tiled_copy_tv entry points shown in the reproducer, then run the provided CuTe DSL script to reproduce the abort. The work is done when the incorrect make_tiled_copy call reports an error instead of terminating with a core dump, while the intended make_tiled_copy_tv path continues to work.