[CodePartition] Optimize the placement of consumer releases

[htyu]

Fix the inefficient placement of consumer releases when producer and consumer are not in the same scope. For example, given

Q = tl.load
for (..) 
   K = tl.load
   QK = dot(Q, K)
   ...
tl.store

Previously the consumer release corresponding to Q was placed after the store. With the current fix the release would go right after the for loop.

TORCH_CUDA_ARCH_LIST=9.0a python run.py --op flash_attention --only triton_tutorial_flash_v2_tma_ws,triton_tutorial_flash_v2_tma_ws_persistent,triton_tutorial_flash_v2 --num-inputs 1 --seq-len 10 --metrics tflops --batch 1024 --n-heads 4 --d-head 128 --cudagraph

Before:

(Batch, Heads, SeqLen, Dhead)    triton_tutorial_flash_v2_tma_ws-tflops    triton_tutorial_flash_v2_tma_ws_persistent-tflops    triton_tutorial_flash_v2-tflops
-------------------------------  ----------------------------------------  ---------------------------------------------------  ---------------------------------
         (1024, 4, 1024, 128)                                   393.141                                              400.046                            366.498

After:
  (Batch, Heads, SeqLen, Dhead)    triton_tutorial_flash_v2_tma_ws-tflops    triton_tutorial_flash_v2_tma_ws_persistent-tflops    triton_tutorial_flash_v2-tflops
-------------------------------  ----------------------------------------  ---------------------------------------------------  ---------------------------------
           (1024, 4, 1024, 128)                                    396.43                                              422.847                            363.753

[manman-ren]

Great perf win!