ci: add perf and density benchmarks for Linux and Windows

[danbugs]

Summary

• Adds a new benchmarks.yml workflow that measures pyhl performance and VM density on both Linux and Windows
• Builds the python-agent-driver image once on Linux and shares it via artifact upload (same pattern as test-examples.yml)
• Reports all results as ::notice:: annotations in the GitHub Actions summary

What's measured

Performance (both platforms):

• Hello world — print(42), 15 runs
• Pandas — DataFrame(np.random.randn(1000,4)).describe(), 10 runs
• Reports median, avg, min, max for each

Density (both platforms):

• 5 concurrent VMs running sum(range(500M)) (~15s compute per VM)
• Linux: Private_Dirty from /proc/$pid/smaps_rollup
• Windows: PrivateMemorySize64 from the process object

Snapshot size (both platforms):

• Apparent size vs actual disk usage
• Linux: stat blocks; Windows: GetCompressedFileSizeW