Do not pool failed CUDA streams

[fallintoplace]

Summary

When CUDA stream creation fails, stream_provider still engages own_stream_ with the null handle. Its destructor then returns that handle to the pool, and a later borrower can receive it with cudaSuccess.

• Keep own_stream_ disengaged when stream acquisition fails.
• Add a CUDA API interception regression test in an isolated executable.
• Avoid returning failed handles to the stream pool.

Tests

• cmake --build build --target test.nvexec.stream_pool --parallel 4
• ./build/test/nvexec/test.nvexec.stream_pool
• cmake --build build --target test.exec test.stdexec --parallel 4
• ./build/test/test.stdexec
• ./build/test/exec/test.exec

[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 d8a004c