Faster adaptive_p sampling

[ikawrakow]

This PR further optimizes adaptive_p sampling compared to PR #1161. For more context, see the discussion there.

To actually measure the time spent in the adaptive_p sampler, one needs to add up the time spent in all of its functions, not just the final sampling time, which is fast. This is done in this PR and also in #1161, but I also modified the main branch (not pushed here) to be able to compare.

Here the results of a quick experiment with Qwne3-30B-A3B-Q8_0, adaptive_p enabled, prompt "Give me an extended summary of the history of Bulgaria". We see a massive improvement between the main branch and #1161 (~17X), and an additional speedup of 2.4X in this PR.

Main branch

llama_print_timings:        load time =    3036.16 ms
llama_print_timings:      sample time =   26234.31 ms /  2617 runs   (   10.02 ms per token,    99.75 tokens per second)
llama_print_timings: prompt eval time =      69.45 ms /    19 tokens (    3.66 ms per token,   273.59 tokens per second)
llama_print_timings:        eval time =   16612.06 ms /  2616 runs   (    6.35 ms per token,   157.48 tokens per second)
llama_print_timings:       total time =   56298.48 ms /  2635 tokens

PR #1161

llama_print_timings:        load time =    2938.55 ms
llama_print_timings:      sample time =    1524.91 ms /  2617 runs   (    0.58 ms per token,  1716.16 tokens per second)
llama_print_timings: prompt eval time =      70.19 ms /    19 tokens (    3.69 ms per token,   270.69 tokens per second)
llama_print_timings:        eval time =   16584.15 ms /  2616 runs   (    6.34 ms per token,   157.74 tokens per second)
llama_print_timings:       total time =   33213.94 ms /  2635 tokens

This PR

ik/adaptive_p_2
llama_print_timings:        load time =    2954.32 ms
llama_print_timings:      sample time =     627.39 ms /  2667 runs   (    0.24 ms per token,  4250.91 tokens per second)
llama_print_timings: prompt eval time =      69.66 ms /    19 tokens (    3.67 ms per token,   272.75 tokens per second)
llama_print_timings:        eval time =   16910.88 ms /  2666 runs   (    6.34 ms per token,   157.65 tokens per second)
llama_print_timings:       total time =   38529.50 ms /  2685 tokens

[ikawrakow]

OK, I added AVX2 implementation of the probabilities. With that I get for the above test case

llama_print_timings:        load time =    2985.43 ms
llama_print_timings:      sample time =     476.95 ms /  2667 runs   (    0.18 ms per token,  5591.72 tokens per second)
llama_print_timings: prompt eval time =      70.72 ms /    19 tokens (    3.72 ms per token,   268.67 tokens per second)
llama_print_timings:        eval time =   16940.82 ms /  2666 runs   (    6.35 ms per token,   157.37 tokens per second)
llama_print_timings:       total time =   33074.71 ms /  2685 tokens

This is, somewhat disappointingly, only 0.58/0.18 = 3.2 times faster than #1161. But if we take into account the time spent in other samplers (0.07 ms in this example), the speedup compared to #1161 becomes (0.58 - 0.07)/(0.18 - 0.07) = 4.6 times. If I campare to the previous main branch, speedup is (10.02 - 0.07)/(0.18 - 0.07) = 90.5 times!