optimize paged attention on triton3

[grimoire]

triton3 has move the cuda fast math location.
This PR support fast expf in paged attention with triton3.0.

Note

None-cuda backend end might not work.

The fill kv kernel and attention is updated so we can change kv layout in the future.

[grimoire]

python mdeploy/benchmark/profile_generation.py \
   internlm2_5-7b-chat-1m/ \
    --tp 4 \
    -c 1 \
    -ct 1 \
    -pt 1000000 \
    --session-len 1048576 \
    --backend pytorch -w 0 -tr 1

--------------------------------------------------
total time: 683.11s
concurrency: 1, test_round: 1
input_tokens: 1000000, output_tokens: 1
first_token latency(min, max, ave): 679.787s, 679.787s, 679.787s
total_token latency(min, max, ave): 679.787s, 679.787s, 679.787s
token_latency percentiles(50%,75%,95%,99%)(s): [679.787, 679.787, 679.787, 679.787]
throughput(output): 0.0 token/s
throughput(total): 1463.9 token/s
--------------------------------------------------

[AllentDan]

May update UT for the new layout type.

[RunningLeon]

flash_attn-2.6.3+cu118torch2.3cxx11abiFALSE-cp38-cp38-linux_x86_64.whl

[AllentDan]

LGTM

[lvhan028]

what does rep mean?

[grimoire]

repeat benchmark for autotuning.

[lvhan028]

How long does this setting warmup=10, rep=25 take?

[grimoire]

each tuning config take 34~35ms

[RunningLeon]

LGTM