[tetsuo-cpp]

GPU test runs fail at random when Vast.ai rents a host whose driver is older than the CUDA 12.4 container image. Every test then fails at cuInit with CUDA error 804. This PR stops renting those hosts.

Root cause

I confirmed this by renting offer 45600079, the host from failed run 40082840322, and running a trivial kernel on it:

NVIDIA GeForce RTX 3060, 535.154.05
	libcuda.so.1 => /usr/local/cuda-12.4/compat/libcuda.so.1
RUN FAILED: SSH command failed (rc=1):
CUDA error at /tmp/warpforth-runner.cpp:289: forward compatibility was attempted on non supported HW

The image's library search order puts its forward-compatibility libcuda (550.54.15) ahead of the host driver's. On a host with an older driver, CUDA tries forward-compatibility mode, which only datacenter GPUs support. A consumer GPU like this RTX 3060 fails with error 804. The offer search never checked which CUDA version the host driver supports.

Changes

• CUDA_VERSION = "12.4" is now one constant. Both the image tag and a new cuda_max_good>=12.4 filter in the offer search come from it.
• The launch log records the GPU model, driver version and cuda_max_good, so a future host problem can be diagnosed from CI logs.
• A new unit test checks that the offer query and the image agree on the CUDA version. It is committed before the fix.

I didn't make the runner treat error 804 as an infrastructure failure, which the issue listed as optional. With the filter in place we no longer rent those hosts.

Verification

• uv run pytest gpu_test/test_vast_session.py gpu_test/test_compiler.py passes 11 tests. The new test fails on the parent commit.
• A live Vast.ai search with the new query, with no rental, returns 10 offers with cuda_max_good of 12.6 or higher. Offer 45600079 is not among them.
• The gpu-test label on this PR runs the full GPU suite.

Fixes #131