Fix CUDA long-context smoke test linking

[Matthley]

Summary

• link ds4_image.o into tests/cuda_long_context_smoke
• resolve ds4_deepseek4_attention_bounds for cuda-regression

Requested by @GiorgioOppo in #952.

Test plan

• make clean && make CUDA_ARCH=sm_121 cuda-regression
• make -j4 cuda-spark
• DGX Spark GB10, CUDA 13.0, sm_121a

[adamlawi]

Independent confirmation on a second DGX Spark (GB10 / sm_121, CUDA 13.0).

Reproduced the failure first, on the current head of #952 (7a5002de) without this patch:

/usr/bin/ld: ds4_cuda.o: in function `ds4_gpu_attention_visual_mixed_batch_heads_tensor':
ds4_cuda.cu:19444: undefined reference to `ds4_deepseek4_attention_bounds'
collect2: error: ld returned 1 exit status
make: *** [Makefile:622: tests/cuda_long_context_smoke] Error 1

With the one-line change applied, it links and the test passes:

ds4: CUDA backend initialized on NVIDIA GB10 (sm_121) dev=0
cuda-regression: top-k n_comp=32768 n_tokens=32 elapsed=0.000s
cuda long-context regression: OK

Worth noting for anyone else building on CUDA: make cuda-spark is unaffected — ds4_image.o is already in CORE_OBJS, so only the tests/cuda_long_context_smoke link line is short of it. That is presumably why this went unnoticed; we have built and shipped several CUDA production binaries from this tree without ever linking that target.

[GiorgioOppo]

merged with some conflicts resolved