(temp workaround for Triton bug) (ROCm#306)

Triton's autotuner had a bug (fixed in
triton-lang/triton@184fb53 )
that we happen to trigger - this PR is a temp. workaround before we
patch/update Triton.

Author: ilia-cher

vllm/attention/ops/triton_flash_attention.py

@@ -367,12 +367,12 @@ def attn_fwd(
     IS_CAUSAL: tl.constexpr,
     BLOCK_M: tl.constexpr,
     BLOCK_DMODEL: tl.constexpr,
+    USE_FP8: tl.constexpr,
     BLOCK_N: tl.constexpr,
     PRE_LOAD_V: tl.constexpr,
     BIAS_TYPE: tl.constexpr,
     ENABLE_DROPOUT: tl.constexpr,
     RETURN_ENCODED_SOFTMAX: tl.constexpr,
-    USE_FP8: tl.constexpr,
 ):
     start_m = tl.program_id(0)
     off_h_q = tl.program_id(1)