everyvoice/tests/test_cli.test_synthesize is broken

the test_synthesize generates fake checkpoints for text-to-spec and spec-to-wav models and then runs synthesize on them. It shouldn't produce good audio with untrained weights, but it should produce something. the tests work on gpsc-c and on my mac, but fail consistently on github actions. I've excluded the test from https://github.com/roedoejet/EveryVoice/pull/252 to avoid the PR getting too stale, but we should fix this to improve testing on an otherwise untested part of the code.

Start with everyvoice/tests/test_cli.test_synthesize and compare its GitHub Actions behavior with the local environments described in the issue; inspect how the test was excluded from PR #252. Done means the test runs in GitHub Actions with fake checkpoints and confirms that synthesize produces output without requiring good audio.

Written by the indexing model from the issue text.