Merge upstream DeepGEMM through PR #432 - #15

Merge upstream DeepGEMM through 66081d4c9c7d7c44f13fea402e5b622aa0f409c2 (PR #432), including the July release and tensormap race fix. The merge retains all six fork commits and ports the fork's training paths onto upstream's DeepJIT runtime, buffer layout, and scheduler.

• Preserve GeGLU, BF16/MXFP4 backward kernels, activation saves across ring wraps, route-weight/combine modes, strided PSUM scales, and configurable SM headroom.
• Expose upstream shared-expert inference and FP8 expert weights while retaining the fork's existing Python positional arguments and native buffer-slicer tuple formats.
• Adapt backward output storage and grouped-GEMM K tiles to upstream's interfaces; keep ordinary inference free of host-synchronized config collectives for CUDA graph capture.
• Retain upstream inference tests in test_mega_moe.py and move the fork's training suite to test_mega_moe_training.py. Add CPU API regressions and a reproducible offline CUDA compilation suite.

Validation passed with GCC 13 and the CUDA/PyTorch development container:

• C++ extension build and Python packaging.
• Six CPU compatibility/training regression tests.
• 26 offline SM100 CUDA variants, including shared experts, FP8/FP8, training saves, 32/240-row tiles, backward, fused weight gradients, and strided PSUM packing.
• Python syntax and shell syntax checks; both the previous fork tip and upstream target are ancestors of the merge commit.

Draft because this host has no working NVIDIA driver: GPU numerical correctness, distributed synchronization, and performance tests have not run. Shared experts combined with training saves or nondefault route/combine modes are explicitly rejected. Existing raw buffers must be reallocated for the new layout and 1920-token alignment; explicit ring capacities must satisfy the new scheduler's estimate. See docs/fork-upstream-sync.md for compatibility details and validation commands.

Please use a merge commit when landing this sync to retain upstream ancestry.