Unit test goals for model interface

[one-some]

lets not explode everything when poking around in model code! :^) tests should ensure:

• [x] Certain outputs on certain seeds are both deterministic and equal to expected outputs
• [ ] Probabilities add up to 1
• [ ] Samplers are no-ops when invalid/"disable" parameters are used
• [x] Correct amount of tokens are generated
• [ ] Token callbacks work (on models where its applicable)
• [x] GPU load/gen works
• [ ] CPU load/gen works
• [ ] GPU x CPU split load/gen works
• [ ] GPU x GPU split load/gen works
• [ ] TPU load/gen works (maybe not this one, not sure how we'd autotest that)
• [ ] biasing and other logit manipulation works (measure logit from known seed or whatever and measure difference with bias)
• [ ] Core gen (maybe, not exactly sure how to test this)
• [ ] Maybe keep track of inference time to watch for regressions or improvements
• [ ] api works

notes:

• obviously we should use an itty bitty model, maybe pythia-70m?
• not sure how hardware differences account into determinism, keep it in mind