set paged attn block size as a env parameter

[jiqing-feng]

Based on our latest tests, the #1065 has 20%+ speed-up for 1st token but will cause 2nd token latency regression, especially on short prompt input. We can set the block size as a ENV parameter and default to 16 to eliminate the 2nd latency regression mostly, and it also keeps most of the acceleration for 1st token.

Overall, with #1065 and this PR, we got the following performance speed-up table:
input len: 512; output len: 128; bs: 2; num_beams: 2

|  speed-up   |   gpt2-xl   | llama3.1-8B |
|  1st token  |    1.37x    |    1.29x    |
|  2nd token  |    0.98x    |    0.97x    |

[HuggingFaceDocBuilderDev]

The docs for this PR live here. All of your documentation changes will be reflected on that endpoint. The docs are available until 30 days after the last update.

[IlyasMoutawwakil]

LGTM, thanks for the perf numbers !