Online simulation for small acceptance rates

[keyboardAnt]

I've been working toward generalizing the heatmap (Figure 1 in the paper) so that we can plug in online simulations instead of only offline. Calculating the heatmap requires running simulations for a grid of configurations. Unfortunately, our current online simulation implementation doesn't support small acceptance rates. This issue not only prevents us from creating heatmaps but also reveals a bug 🥲

Thanks to edge case 1 described below, generating S tokens shouldn't take much longer than S * (failure_cost + c + wait_for_pipe). However, our current implementation ignores the edge case, causing the latency to scale with 1/acceptance_rate. For example, generating S=77 tokens with acceptance_rate=0.01 requires approximately 7700 iterations instead of <= 77. For reference, I added tests covering the issue (see the two skipped tests in tests/online/test_simul.py, named test_correct_token_count_per_iteration and test_duration)

The edge cases (from private email correspondence I sent on July 15):

Two edge cases of DSI boost its speedup but are overlooked in our online simulations. The two suggested changes below will improve the speedups reported in the paper (affecting Table 1 and Figure 1).

1. Accept an extra token if the target rejects a draft or if the extra token is the last.
2. Simulate an immediate validation of the extra token by terminating the corresponding speculating iteration with probability 1 - acceptance_rate.

For more context, please see my PR comment.

[keyboardAnt]

Since test_correct_token_count_per_iteration was a reimplementation of the core logic and we are editing the core logic as we go, I removed this test in #43. Instead, I added test_num_of_fix_history.