[BUG] [cuteDSL] mark_compact_shape_dynamic fails to pass shape args to JIT kernel (resolves to 0 at runtime)

Issue Description

Hi, I've encountered a bug where dynamic tensor shapes marked with mark_compact_shape_dynamic are not correctly marshalled to the device kernel. When using static shapes, the kernel receives the correct values.

Behavior:

• With mark_layout_dynamic & mark_compact_shape_dynamic: Inside the @cute.kernel, the dynamic dimensions resolve to 0 and strides contain garbage values, even when passing a valid concrete runtime tensor.
• Without these dynamic markers (Static Shape): The kernel correctly receives the shape values (e.g., (128, 128)).

Steps to Reproduce:

1. Create a template GPU tensor.
2. (Bug Case) Mark layout and dimensions as dynamic using mark_layout_dynamic and mark_compact_shape_dynamic.
3. Compile a JIT kernel that launches a device kernel.
4. Pass a concrete runtime tensor.
5. Inspect tensor.shape inside the device kernel via cute.printf.

Reproduction Script:

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

# Device Kernel
@cute.kernel
def device_kernel(tensor: cute.Tensor):
    tidx, _, _ = cute.arch.thread_idx()
    if tidx == 0:
        # BUG: These prints show (0, 0) when using dynamic shape
        # CORRECT: Shows (128, 128) when using static shape
        cute.printf("Device Kernel >> Tensor: {}\n", tensor)
        cute.printf("Device Kernel >> Shape: ({}, {})\n", tensor.shape[0], tensor.shape[1])

# Host Launcher
@cute.jit
def launcher_func(tensor: cute.Tensor):
    # Launch kernel with 1 thread to inspect values
    device_kernel(tensor).launch(grid=(1,1,1), block=(1,1,1))

def main():
    M, N = 128, 128
    
    # 1. Prepare Template (GPU Tensor -> gmem)
    fake_input = torch.zeros(M, N, device='cuda', dtype=torch.bfloat16)
    template = from_dlpack(fake_input, assumed_align=8)
    
    # [BUG REPRODUCTION]
    # Case A (Bug): Dynamic Shape -> Device sees (0, 0)
    # Case B (Working): Static Shape -> Device sees (128, 128)
    # Comment out the following lines to switch to Case B
    template = template.mark_layout_dynamic(leading_dim=1)
    template = template.mark_compact_shape_dynamic(
        mode=1,
        stride_order=(0, 1),
        divisibility=16
    )
    
    print(f"Compiling with template: {template}")
    
    compiled_func = cute.compile(launcher_func, template)
    
    print("\nRunning compiled function...")
    input_real = torch.randn(M, N, device='cuda', dtype=torch.bfloat16)
    input_runtime = from_dlpack(input_real, assumed_align=8)
    
    compiled_func(input_runtime)
    torch.cuda.synchronize()

if __name__ == "__main__":
    main()

Actual Output (Dynamic Shape - Bug):

Device Kernel >> Shape: (0, 0)

Expected Output (Static Shape - Reference):

Device Kernel >> Shape: (128, 128)

Thanks.