[None][fix] Revert "Merge Eagle3 and MTP-eagle one-model workers (#12353)"

[tensorrt-cicd]

This reverts commit 8e5d9e2 (#12353).

Summary

The MTP-eagle / Eagle3 one-model worker merge (#12353) regresses the GB200 2-node disagg perf-sanity gen-only test:

test_perf_sanity.py::test_e2e[disagg_upload-gen_only-gb200_deepseek-v32-fp4_1k1k_con2048_ctx1_dep4_gen1_dep4_eplb0_mtp1_ccb-NIXL]

Failure signature

The GEN server (MTP speculative decoding, num_nextn_predict_layers=1) crashes during executor init while capturing the generation-phase CUDA graphs. All 4 gen ranks abort with exit code 134, so the benchmark never runs:

File "model_engine.py", line 1169, in _run_cuda_graph_warmup
    self._capture_generation_cuda_graphs(resource_manager)
File "model_engine.py", line 1240, in _capture_generation_cuda_graphs
    torch.cuda.synchronize()
torch.AcceleratorError: CUDA error: unspecified launch failure

Bisect

Comparing the last passing post-merge commit 33b0a32 with the first failing one 316430f (12 commits): #12353 is the only change touching the spec-decode / model_engine CUDA-graph capture path. The remaining commits in the range are AutoDeploy, test, and infra changes (none touch deepseek-v32 / MoE / attention / CUDA-graph / spec-decode kernels). The cutlass-dsl version is unchanged across the range.

Reverting restores gen-server startup and unblocks the perf-sanity test.

Summary by CodeRabbit

• Refactor

  • Refactored EAGLE3 one-model speculative decoding to streamline linear-tree processing and remove legacy code paths.
  • Simplified speculative decoding worker constructors and method signatures for improved maintainability.
  • Consolidated speculative metadata handling and removed redundant parameters.

• Chores

  • Cleaned up license header comments and reorganized import statements across speculative decoding modules.
  • Updated configuration field types in MTPDecodingConfig from constrained to standard numeric types.