Clean up modeling_deepseek.py

[hlu1]

• Rename some variables to make the code more readable
• Remove unnecessary if/else branches
• Simplify some of the logics in allreduce fusion

The modeling_deepseek.py files looks mostly clean after this round of cleaning. I kept the allreduce part the same because we need the allreduce op unification PR to land first.

For accuracy test, I ran the examples/pytorch/quickstart_advanced.py script with the R1 model with different configs.

[hlu1]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #2533 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2533 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1815 completed with status: 'FAILURE'

[hlu1]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #2596 [ run ] triggered by Bot

[hlu1]

/bot kill

[tensorrt-cicd]

PR_Github #2601 [ kill ] triggered by Bot

[tensorrt-cicd]

PR_Github #2596 [ run ] completed with state ABORTED

[tensorrt-cicd]

PR_Github #2601 [ kill ] completed with state SUCCESS
Successfully killed previous jobs for commit e192b6e

[hlu1]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #2605 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2605 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1857 completed with status: 'FAILURE'

[hlu1]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #2672 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2672 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1903 completed with status: 'SUCCESS'

[QiJune]

LGTM