sampling: make top_n_sigma no-op at <=0 rather than <0

[DocShotgun]

This changes the behavior of the recently-added top_n_sigma sampler to a short-circuit no-op state at values <= 0 rather than < 0. The rationale for this change is as follows:

• The current behavior of top_n_sigma == 0 is redundant as it is a more roundabout way to achieve greedy decoding, which already has other means of being specified, i.e. top_k == 1
• top_n_sigma == 0 represents no-op rather than greedy decoding in other existing tooling (i.e. text-generation-webui, aphrodite-engine, koboldcpp, YALS), so this would keep the interface consistent for frontend developers

[CISC]

Do you know the rationale for koboldcpp also checking for cur_p->size <= 1?

Well, looking closer I see why, so perhaps add that too?

[CISC]

Ouch, test-sampling fails, can you look into why?

[DocShotgun]

Ouch, test-sampling fails, can you look into why?

I took a look at it, and far as I can tell, it's because this line:

llama.cpp/tests/test-sampling.cpp

Line 363 in 91a86a6

test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {1.0f, 0.0f, 0.0f, 0.0f}, 0.00f);

explicitly checks for top_n_sigma == 0 leading to greedy decoding, and that behavior is changed by this PR.

If I change the test to check for no-op instead, it passes:

test_top_n_sigma({0.1f, 0.2f, 0.3f, 0.4f}, {0.4f, 0.3f, 0.2f, 0.1f}, 0.00f);

[CISC]

If I change the test to check for no-op instead, it passes

Great, can you also add a comment explaining why it was changed?