add WORKERD_ENABLE_ALL_AUTOGATES env variable

[anonrig]

Ref: #4089

Adds an environment variable called "WORKERD_ENABLE_ALL_AUTOGATES" which enables all autogates regardless of what's set in the config. This is beneficial for testing.

[jasnell]

I'm not entirely sure this is the right approach but could be convinced. The key reason is that this potentially misses the various combinations of some-autogates-on-some-autogates-off that could expose bugs and might be missed by this... For instance, the code might work if both autogates A and B are off or both are on, but does it work if only one of the two are on and the other is off? What I think might be better is a bazel configuration that automatically sets up variations of each test for each combination of autogates. That is, if we have three active autogates, then we would have eight test permutations covering each of the various combinations.

[danlapid]

I'm not entirely sure this is the right approach but could be convinced. The key reason is that this potentially misses the various combinations of some-autogates-on-some-autogates-off that could expose bugs and might be missed by this... For instance, the code might work if both autogates A and B are off or both are on, but does it work if only one of the two are on and the other is off? What I think might be better is a bazel configuration that automatically sets up variations of each test for each combination of autogates. That is, if we have three active autogates, then we would have eight test permutations covering each of the various combinations.

This PR mimics what we have internally, we lack the same detail with our internal tests.
It's a balance between coverage and test times.
It is reasonable to want better coverage but I think the request is out of scope for this PR.
This mimics correctly what we have internally and I think it is reasonable.

[kentonv]

Should we consider using an environment variable instead, so that we can also run KJ tests with all-autogates?

That is, if we have three active autogates, then we would have eight test permutations covering each of the various combinations.

Unfortunately this doesn't work because it doesn't scale. Even 8x is already probably too much, but in the internal codebase we have 21 autogates defined currently, in addition to the 3 in workerd, that's 24 autogates, so we'd be running each test 16 million times. Granted many of them should probably be cleaned up but still...

I think running one with all off and once with all on is a reasonable compromise. Autogates rarely interact with each other in practice.

[anonrig]

Should we consider using an environment variable instead, so that we can also run KJ tests with all-autogates?

I don't have any strong opinion. I'm fine with replacing the cli flag with an environment variable.