tests: try to make s2n_mem_usage_test more useful

[lrstewart]

Release Summary:

Resolved issues:

resolves #5121 (comment)

Description of changes:

I always dread touching s2n_mem_usage because changing the expected values always breaks the test SOMEWHERE. I've also had the test fail locally even though I updated MEM_PER_CONNECTION to the value the test told me to use, just because of the complicated math. So this PR:

1. Skips the test almost everywhere. Only enable the test in specific CI builds.
2. Simplifies the math. I convert the measured vm memory to memory per connection, rather than trying to go the other direction.
3. Allows less variance. We should really know if memory increases or decreases by more than 5%. That's not trivial.
4. Don't rely on TEST_DEBUG_PRINT. We should just always print the output on failure.

Testing:

Example failure output from running it locally:

Actual KB per connection: 45
This is a -11.76% change

VmData initial:                  655360
VmData after allocations:      16117760
VmData after handshakes:       24354816
VmData after free handshake:   24354816
VmData after release:          24354816
Number of connections used:         252

By submitting this pull request, I confirm that my contribution is made under the terms of the Apache 2.0 license.

[dougch]

good riddance to this bit.

[lrstewart]

Thinking about it, should this be even stricter? 3%? Or should any change to the final "KB" number require an update to the test-- so 0%? That would still let changes <1KB through.

[goatgoose]

1KB seems like kind of a lot, right? I feel like 0% would be enough variance. Not sure though.

[lrstewart]

Let's do 0 and we can always add some wiggle room later if we need it.

[goatgoose]

1KB seems like kind of a lot, right? I feel like 0% would be enough variance. Not sure though.