Tests may fail when run via bazel test

[brandjon]

After the fix to tests for #98, tests may fail when run directly via bazel test as opposed to through the runner script run_tests.sh. I see two modes of failure:

1. When --incompatible_use_python_toolchains is disabled, //tests:version_test may fail if /usr/bin/env python is a Python 2 interpreter, due to py_binary doesn't automatically pick a Python runtime matching the requested version bazelbuild/bazel#4815.

2. When --incompatible_use_python_toolchains is enabled, the default toolchain's py[2|3]wrapper.sh script is added to runfiles, causing brittle assertions in many tests to fail. This doesn't appear under run_tests.sh because it has its own custom toolchain that adds no runfiles.

This isn't an immediate problem because CI still passes, but it is annoying for the developer workflow. For 1), I'd just wait for the toolchain flag flip. For 2), the test assertions should be made more robust. In the meantime, use run_tests.sh instead of bazel test //....

[brandjon]

Turns out it is a problem for Bazel downstream CI, just not subpar's own CI.