Write simple additional performance tests

[spigo900]

Addresses part of #45.

cc @benlebrun

[spigo900]

@benlebrun As it stands:

1. I run the tests using make benchmark to run them all, or pytest perf_tests/test_inference -k test_name_substring to run a single test.
2. The long sequences test should be short enough -- it ran in 1 minute 45 seconds locally, total, when I ran only that test.
3. The permissive grammar test takes ~5 minutes as it stands, so I would shorten the token limit from 100 tokens to 40 tokens I think and adjust from there.
4. I don't know where the "many particles" benchmark stands.
5. The benchmarks seem to take longer than the benchmark printout reports, probably because of time spent loading the model. It might be possible to fix these by putting model loading into a fixture, but I don't know if there is a nice way to do that because each test needs to use its own grammar in the inference setup.

So, the tests are in place but the parameters still need adjusting, and we might be able to increase the test parameters if we can factor out the model loading/inference setup somehow.