Free device allocation when copy fails

[fallintoplace]

Problem

device_allocate calls cudaMalloc and then copies the initial value with cudaMemcpy. If that copy returns an error, the function returns an empty pointer, but the successful device allocation has no owner and is leaked.

The exception path already tried to free the allocation, which made the missing normal-error cleanup easier to miss.

Fix

• Give the allocation a device_ptr_t owner immediately after cudaMalloc succeeds.
• Return that owner on success.
• Let the owner release the allocation when cudaMemcpy fails or host construction throws.

Test

• Added a focused test that forces cudaMemcpy to fail and counts the matching cudaFree call.
• clang-format-21 --dry-run --Werror
• CMake configure and build with CUDA disabled

[copy-pr-bot]

This pull request requires additional validation before any workflows can run on NVIDIA's runners.

Pull request vetters can view their responsibilities here.

Contributors can view more details about this message here.

[ericniebler]

/ok to test 7779c5d