Tune BF16 wgrad and fuse DeepEP metadata scale packing

[A-nnonymous]

Summary

• tunes the Blackwell BF16 wgrad default path and adds coverage for the varlen-K BF16 wgrad path
• adds fused DeepEP topk metadata + raw-scale ISA packing support with fallback to the existing metadata path
• plumbs prepacked activation scales into the Sonic up-projection path so the standalone raw-scale gather can be skipped when the fused path is active
• extends the metadata NCU harness with production A35B shape and pack-mode switches
• adds a PR-facing metadata/prelaunch bubble breakdown and experiment/test matrix in artifacts/deepep_sonic_metadata_gap_breakdown_20260709.md

Why

The latest A35B timeline shows Sonic compute is no longer the primary blocker after BF16 wgrad tuning. The remaining DeepEP -> Sonic up-projection path has a stable host-side bubble: the old sequence still shows _gather_raw_scales_1x32_to_isa_kernel, followed by about 485 us p50 before the Sonic up GEMM launch. Same-launch-thread NVTX attributes stable portions of that bubble to view_dtype, empty, transpose, allocator work, and generic CUTE/QuACK wrapper setup.

The fused metadata+scale path is the first step: it removes the separate host transition into raw-scale packing and gives the next optimizer a clear test/perf matrix for eliminating the remaining wrapper/carrier costs without changing numerics.

Validation

• CUDA_VISIBLE_DEVICES=0 python -m pytest tests/ops/test_deepep_topk_metadata.py::TestCudaScalePacking -q --tb=short -> 9 passed
• CUDA_VISIBLE_DEVICES=0 python -m pytest tests/ops/test_deepep_topk_metadata.py -q --tb=short -> 74 passed
• CUDA_VISIBLE_DEVICES=0 python tests/ops/ncu_deepep_topk.py --config prod_a35b --with-scales --repeat 10 -> avg_gpu_ms=0.356854, packed_scales_shape=(1, 208044032)

Notes For Reviewers

• This PR does not disable or modify the EP barrier.
• The current A35B path is DeepEP, not MoE allgather.
• The new analysis document distinguishes measured facts from proposed follow-up experiments and lists the required bit-exact tests for each follow-up.
• Untracked local NCU .ncu-rep artifacts were intentionally not committed.