fix: fix cuda graph padding for spec decoding

[lfr-0531]

Description

Root cause:

• Previously, we used if request.py_batch_idx is None to identify dummy requests when enabling overlap scheduler. However, after the CUDA Graph padding changes in [TRTLLM-5516] perf: replicate dummy request for cuda graph padding #4729, the same dummy request is reused across all model forward passes, causing request.py_batch_idx to be non-None. This leads to an error.

In this PR:

• Change to use if request.is_dummy to identify dummy requests.
• Move the dummy extend requests to the end of extend_requests to align with the logic for those requests in generation_requests.
• Add a test for CUDA Graph padding + spec decoding.

Test Coverage

• Add new tests:
  • accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales_cuda_graph_padding[mtp_nextn=2]
  • accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales_cuda_graph_padding_4gpus[attention_dp=True-mtp_nextn=0]
  • accuracy/test_llm_api_pytorch.py::TestDeepSeekV3Lite::test_fp8_block_scales_cuda_graph_padding_4gpus[attention_dp=True-mtp_nextn=2]

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #7289 [ run ] triggered by Bot

[QiJune]

LGTM

[tensorrt-cicd]

PR_Github #7289 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5281 completed with status: 'FAILURE'

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #7307 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7307 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5295 completed with status: 'FAILURE'

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #7344 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7344 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5323 completed with status: 'FAILURE'

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #7413 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7413 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5378 completed with status: 'FAILURE'

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #7441 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7441 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5402 completed with status: 'FAILURE'

[lfr-0531]

/bot run

[tensorrt-cicd]

PR_Github #7469 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7469 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5422 completed with status: 'FAILURE'

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #7542 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7542 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5472 completed with status: 'FAILURE'

[lfr-0531]

/bot kill

[tensorrt-cicd]

PR_Github #7691 [ kill ] triggered by Bot

[tensorrt-cicd]

PR_Github #7688 [ run ] completed with state ABORTED

[tensorrt-cicd]

PR_Github #7691 [ kill ] completed with state SUCCESS
Successfully killed previous jobs for commit 08fa91c

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #7699 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7699 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5585 completed with status: 'FAILURE'

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #7773 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7773 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5629 completed with status: 'FAILURE'

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #7831 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7831 [ run ] completed with state FAILURE

[lfr-0531]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #7849 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #7849 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #5668 completed with status: 'SUCCESS'
Pipeline passed with automatic retried tests. Check the rerun report for details.