test: Refactor cpu metrics tests to make L0_metrics more stable

[rmccorm4]

The previous implementation of tests for the CPU/RAM related metrics were:

1. Unstable on the check for nv_cpu_utilization because there was no work going on in the background, making duplicate values appear fairly frequently.
2. Written in bash

This PR:

1. Adds server inference requests in the background to keep the server busy and show some CPU utilization variation during metric observation
2. Moves to python for more flexibility and ease of testing
3. Update the newer styled tests (pinned_memory_metrics, cpu_metrics) to use the new pytest style that generates nice reports on the gitlab side for easier regression detection at the unit test level. I also noticed the use of check_test_results without using the tu.ResultCollector for the pinned memory metrics tests section - so it wasn't actually checking the right test results file anyway, since none was being generated. I moved this one to the pytest style to be consistent with the similar cpu metrics test.
4. Make some windows-friendly changes to the L0_metrics tests based on Francesco's review comments like use of kill_server and TRITONSERVER_IPADDR.
5. Reduces 4GPU CI runner requirement to 2GPU requirement by only checking 2-gpus instead of 3-gpus in the gpu metrics section. I think the test is equivalent regardless, but there are significantly more 2gpu runners available in CI than 4gpu runners to cut down on job queue times. I'll push a gitlab-side PR to update accordingly, but it will still work as-is with these changes.

Average results of the PR from testing so far typically makes about 20 observations (1 per second for 20 second test), and finds no duplicate values. However, keeping the duplicate tolerance will help stabilize in the event we occasionally find some:

[fpetrini15]

TmLg! Great work Ryan! Look forward to seeing the greener pipelines 🚀

And thanks for fixing things up to be Windows CI friendly!

[krishung5]

LGTM!

[fpetrini15]

LGTM, great work!