[Triton] DSV4 fusions phase 1

[k50112113]

This PR includes:

1. q_per_head_norm + qk_rope + swa_write -> fused_reduce_q_norm_qk_rope_swa_write (ATOM_V4_USE_TRITON_FUSION)
2. clamp + silu + mul + quant -> fused_clamp_act_mul_fp8_group_quant (ATOM_V4_USE_TRITON_FUSION)
3. enabled ATOM_ENABLE_DS_QKNORM_QUANT_FUSION in DSV4 (the API is imported from deepseek_v2.py)
4. updates: fuse the upcast of combined = self.wkv_gate(x, otype=torch.bfloat16) into _fused_compress_attn_kernel
5. neg1 = positions.new_full((), -1) at atom/model_ops/attentions/deepseek_v4_attn.py that prevent host-side sync, 5% improvement end-to-end

This PR requires ROCm/aiter#3057

baseline lm_eval

local-completions ({'model': '/data/deepseek-ai/DeepSeek-V4-Flash', 'base_url': 'http://localhost:8000/v1/completions', 'num_concurrent': 16, 'max_retries': 2, 'tokenized_requests': False}), gen_kwargs: ({}), limit: 100.0, num_fewshot: 3, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  | 0.97|±  |0.0171|
|     |       |strict-match    |     3|exact_match|↑  | 0.97|±  |0.0171|

fused lm_eval

local-completions ({'model': '/data/deepseek-ai/DeepSeek-V4-Flash', 'base_url': 'http://localhost:8000/v1/completions', 'num_concurrent': 16, 'max_retries': 2, 'tokenized_requests': False}), gen_kwargs: ({}), limit: 100.0, num_fewshot: 3, batch_size: 1
|Tasks|Version|     Filter     |n-shot|  Metric   |   |Value|   |Stderr|
|-----|------:|----------------|-----:|-----------|---|----:|---|-----:|
|gsm8k|      3|flexible-extract|     3|exact_match|↑  | 0.97|±  |0.0171|
|     |       |strict-match    |     3|exact_match|↑  | 0.97|±  |0.0171|

so far I am seeing ~10% runtime reduction in the trace in decode and not runtime difference in prefill

baseline decode

fused decode

in end-to-end performance, I saw 8% performance uplift, but ~5% comes from item 5), all other optimization, 1) ~ 4), should be around ~10% of improvement yet we only saw 3%, there may be other things other than 5) that is blocking the performance.

Baseline

Maximum request concurrency: 64
============ Serving Benchmark Result ============
Successful requests:                     512       
Benchmark duration (s):                  210.15    
Total input tokens:                      524288    
Total generated tokens:                  524288    
Request throughput (req/s):              2.44      
Output token throughput (tok/s):         2494.83   
Total Token throughput (tok/s):          4989.66   
---------------Time to First Token----------------
Mean TTFT (ms):                          1205.41   
Median TTFT (ms):                        1251.45   
P99 TTFT (ms):                           2057.02   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          24.49     
Median TPOT (ms):                        24.49     
P99 TPOT (ms):                           25.49     
---------------Inter-token Latency----------------
Mean ITL (ms):                           24.47     
Median ITL (ms):                         23.97     
P99 ITL (ms):                            26.08     
==================================================

With all the optimization in this PR

Maximum request concurrency: 64
============ Serving Benchmark Result ============
Successful requests:                     512       
Benchmark duration (s):                  193.79    
Total input tokens:                      524288    
Total generated tokens:                  524288    
Request throughput (req/s):              2.64      
Output token throughput (tok/s):         2705.39   
Total Token throughput (tok/s):          5410.78   
---------------Time to First Token----------------
Mean TTFT (ms):                          1060.13   
Median TTFT (ms):                        984.96    
P99 TTFT (ms):                           1554.14   
-----Time per Output Token (excl. 1st token)------
Mean TPOT (ms):                          22.64     
Median TPOT (ms):                        22.55     
P99 TPOT (ms):                           23.45     
---------------Inter-token Latency----------------
Mean ITL (ms):                           22.62     
Median ITL (ms):                         22.10     
P99 ITL (ms):                            24.22     
==================================================

[k50112113]

updates: fuse the upcast of combined = self.wkv_gate(x, otype=torch.bfloat16) into _fused_compress_attn_kernel

[k50112113]

updates: neg1 = positions.new_full((), -1) that prevent host-side sync, 5% improvement end-to-end.