[fix] Fix a few issues with EAGLE3 in PyTorch backend

[mikeiovine]

• Add a few more explanatory comments
• Clean up draft model hidden state saving logic.
• Add eagle3 support to flashinfer (includes CUDA graph support). Flashinfer is still useful for debugging; it helped me identify some issues in the warmup phase.
• Fix an illegal memory access related to maximum sequence length in warmup phase. This currently does not cause crashes on the TRTLLM attention path (but it's still UB!)

[mikeiovine]

/bot run --disable-fail-fast

[tensorrt-cicd]

PR_Github #2791 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #2791 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #1974 completed with status: 'FAILURE'

[mikeiovine]

/bot run

[tensorrt-cicd]

PR_Github #3333 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #3333 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #2325 completed with status: 'FAILURE'

[lfr-0531]

Adding @yuxianq for vis of this PR. There are some changes about the flashinfer backend.

[mikeiovine]

/bot run

[tensorrt-cicd]

PR_Github #3623 [ run ] triggered by Bot

[tensorrt-cicd]

PR_Github #3623 [ run ] completed with state SUCCESS
/LLM/main/L0_MergeRequest_PR pipeline #2559 completed with status: 'SUCCESS'

[mikeiovine]

Will followup with @yuxianq offline if more changes are requested. For now I think the functional changes to the flashinfer backend are small enough to just land this (and they're pretty much limited to the speculation path only). The bulk of this PR is bug fixing.