da3_streaming.py Floating point exception

```
~/ccode/Depth-Anything-3/da3_streaming$ python da3_streaming.py --image_dir ./input50
Sim3solve of C++ Version failed, Will using Python Version.
[WARN ] Dependency `gsplat` is required for rendering 3DGS. Install via: pip install git+https://github.com/nerfstudio-project/gsplat.git@0b4dddf04cb687367602c01196913cde6a743d70
The exp will be saved under dir: ./exps/._input50/2026-02-25-11-12-44
config yaml file has been copied to: ./exps/._input50/2026-02-25-11-12-44/base_config.yaml
Loading model...
[INFO ] using SwiGLU layer as FFN
[INFO ] using MLP layer as FFN
Using cache found in /home/chenxiang/.cache/torch/hub/facebookresearch_dinov2_main
Model loaded: ./weights/dino_salad.ckpt
init done.
Loading images from ./input50...
Found 5 images
Processing 5 images in 1 chunks of size 120 with 60 overlap
[Progress]: 0/1
Loaded 5 images
[INFO ] Processed Images Done taking 0.2578616142272949 seconds. Shape: torch.Size([5, 3, 280, 504])
[INFO ] Selecting reference view using strategy: middle
*** Aborted at 1771989188 (unix time) try "date -d @1771989188" if you are using GNU date ***
PC: @ 0x7857cd76f137 (/home/chenxiang/tool/miniconda3/envs/da3/lib/python3.10/site-packages/nvidia/cublas/lib/libcublas.so.12+0x16f136)
*** SIGFPE (@0x7857cd76f137) received by PID 297926 (TID 0x78587c319740) from PID 18446744072861708599; stack trace: ***
@ 0x78587c0a1ed3 (/usr/lib/x86_64-linux-gnu/libc.so.6+0xa1ed2)
@ 0x7855e4626868 (/home/chenxiang/tool/miniconda3/envs/da3/lib/python3.10/site-packages/pycolmap/_core.cpython-310-x86_64-linux-gnu.so+0xa26867)
@ 0x78587c045330 (/usr/lib/x86_64-linux-gnu/libc.so.6+0x4532f)
@ 0x7857cd76f137 (/home/chenxiang/tool/miniconda3/envs/da3/lib/python3.10/site-packages/nvidia/cublas/lib/libcublas.so.12+0x16f136)
@ 0x7857cd87e4b3 (/home/chenxiang/tool/miniconda3/envs/da3/lib/python3.10/site-packages/nvidia/cublas/lib/libcublas.so.12+0x27e4b2)
@ 0x7857cd87efec (/home/chenxiang/tool/miniconda3/envs/da3/lib/python3.10/site-packages/nvidia/cublas/lib/libcublas.so.12+0x27efeb)
@ 0x7857cd7293ae cublasSdot_v2
@ 0x7856c687337f void at::cuda::blas::dot(cublasContext*, int, float const*, int, float const*, int, float*)
@ 0x7856c6916f9a at::native::dot_cuda(at::Tensor const&, at::Tensor const&) [clone .localalias]
@ 0x7856c656c890 c10::impl::wrap_kernel_functor_unboxed_, at::Tensor, c10::g@
@ 0x785703145481 at::_ops::dot::call(at::Tensor const&, at::Tensor const&)
@ 0x7857020c7d33 at::autocast::WrapFunction_<(at::autocast::CastPolicy)4, (c10::DeviceType)1, at::Tensor (at::Tensor const&, at::Tensor const&), &at::_ops::dot::call, at::Tensor, c10::guts::typelist::typelist >::call(at::Tensor const&,@
@ 0x785701f94880 c10::impl::wrap_kernel_functor_unboxed_ >, at::Tensor (at::Tensor const@
@ 0x785703145481 at::_ops::dot::call(at::Tensor const&, at::Tensor const&)
@ 0x78571614a6d2 torch::autograd::THPVariable_dot(_object*, _object*, _object*)
@ 0x4fe087 cfunction_call
@ 0x4f77db _PyObject_MakeTpCall.localalias
@ 0x4f3896 _PyEval_EvalFrameDefault
@ 0x4fe4cf _PyFunction_Vectorcall
@ 0x4ee7cf _PyEval_EvalFrameDefault
@ 0x509dde method_vectorcall
@ 0x4f2fd6 _PyEval_EvalFrameDefault
@ 0x509f36 method_vectorcall
@ 0x4f1029 _PyEval_EvalFrameDefault
@ 0x509f36 method_vectorcall
@ 0x4f1029 _PyEval_EvalFrameDefault
@ 0x4f6b2d _PyObject_FastCallDictTstate.localalias
@ 0x5083d0 _PyObject_Call_Prepend
@ 0x5d0c83 slot_tp_call
@ 0x4f77db _PyObject_MakeTpCall.localalias
@ 0x4f3896 _PyEval_EvalFrameDefault
@ 0x4fe4cf _PyFunction_Vectorcall
Floating point exception (core dumped)
```

Reproduce the failure with da3_streaming.py --image_dir ./input50 and review the reference-view selection path after "Selecting reference view using strategy: middle." Trace the interaction between pycolmap and the PyTorch CUDA dot operation shown in the stack trace. Done means the command no longer aborts with SIGFPE and still completes processing the five images.

Written by the indexing model from the issue text.