Only Try a Unique Venv If There Is a Conflict

Currently any failure when installing a benchmark's requirements triggers the creation of a unique venv for that benchmark and a retry of the install. The only failure for which we need to create a unique venv is when there is a conflict with the version of already installed packages. Adjusting the code along those lines would speed things up when their are failures (e.g. wheel cannot build) and produce less noise.

Possible solutions:

• parse the text of the first install attempt to see if it a conflict
  • this is fragile, as well as requiring that we "tee" the output of the "pip install" command
• instead of trying the common venv first, track the installed versions and use the common venv only if the benchmark's requirements are compatible
  • #158 identifies several approaches to tracking installed versions
  • this means we need to add some logic for checking requirements compatibility, though there is probably a good enough library out there (maybe in the stdlib)

Also see https://github.com/python/pyperformance/issues/144#issuecomment-1075822776.

Trace the benchmark requirements-installation flow and review issue #158 for approaches to tracking installed versions. The change is complete when a unique virtual environment is used only for package-version conflicts, while other installation failures are reported without an automatic retry.