Fix CI: work around tox-docker Gateway KeyError on newer Docker

[rudokemper]

Goal

Closes #242. Unblock alerts tox env on CI.

This was not a Python / pip version issue as initially suspected, but rather a discrepancy between tox-docker and ubuntu-2404:current.

What I changed and why

We set TOX_DOCKER_GATEWAY: localhost in .circleci/config.yml.

Newer Docker on ubuntu-2404:current no longer fills top-level NetworkSettings.Gateway, so tox-docker seems to crash with a KeyError. Claude Opus 4.7 did this diagnosis:

tox-docker has a documented escape hatch for exactly this: if TOX_DOCKER_GATEWAY is set, it skips container.attrs["NetworkSettings"]["Gateway"] and uses socket.gethostbyname(gateway) instead. Setting it to localhost resolves to 127.0.0.1, which is what you want on machine-executor anyway (host networking).

How I convinced myself this is right

• Trace lands in tox-docker, not our code; pip freeze is identical to previous green runs on CircleCI.
• alerts tests seems to be passing again in CI.

What I'm not doing here

• Pinning ubuntu-2404 to a dated tag.

LLM use disclosure

Claude Opus 4.7 helped diagnose.

[IamJeffG]

would you mind labeling this as a temp workaround until tox-dev/tox-docker#196 is released?

(if you agree that's true)