Stabilize fork-sensitive nightly tests and AutoSP coverage

[tohtana]

This PR fixes issues with the nightly full test:

• Fork-sensitive tests: Run these tests in the non-forked sequential phase
  • Heavy two-rank NVMe cases contended in the forked parallel phase
  • The CPU-only FlopsProfiler formatting test could inherit initialized CUDA state and fail in a forked subprocess
• AutoSP coverage: TestSDPANodesCompile inspected only the first FX graph produced from a Transformers model, but a compiled model can be split into multiple graphs and the first graph may not contain attention. The test now constructs an FX GraphModule with Q/K/V inputs and one torch.nn.functional.scaled_dot_product_attention call. The separate four-GPU tests still continue to use a tiny Transformers Llama model.
• Transformers version output: Nightly tests install Transformers from Git, but the workflow incorrectly displayed the unused PyPI fallback version (4.50.0) as if it were installed. It now reports the actually installed version and Git commit, shows unused fields as N/A.

[delock]

Looks good to me