…n test

The Phase-3 renormalize block in `grouped_topk_single_group_kernel` called
`warp_sum_f32` (which uses `__shfl_xor_sync(0xffffffff, ...)`) from inside
`if (lane_id < topk)`. With `topk` < 32 (e.g. nemotron-3-nano: topk=6), only
lanes 0..topk-1 reached the intrinsic, but the mask 0xffffffff named all 32
lanes. CUDA spec: every lane named in the mask must execute the intrinsic
at the same site, otherwise the result is undefined.

Empirically the UB returned values from the absent lanes' registers,
producing wrong renormalized weights — 2 of 6 weights per token were
unnormalized (~1.5x too large). The wrong values were tolerated in eager
inference, but under piecewise CUDA graph replay they cascaded into a
downstream OOB that surfaced as IMA at `piecewise_cuda_graph_runner.py:794`
on `TestNvidiaNemotron3Nano30BFP8.test_lm_eval`.

Fix: move the warp_sum out of the divergent `if`, have all 32 lanes
participate, with inactive lanes contributing the additive identity (0).
Output writes remain gated by `if (lane_id < topk)`.

Validated:
- Unit sweep across E in {16..512}, K in {1..8}, N in {1..128}: matches
  reference biased_grouped_topk_impl with max diff < 1e-7.
- 2x H200 e2e: TestNvidiaNemotron3Nano30BFP8.test_lm_eval passes
  (gsm8k strict=0.839, flexible=0.542, both within rtol=0.08).
- Buggy kernel + eager (no graphs) also passes — confirming the kernel
  itself doesn't fault, only the cascade-under-graph-replay does.

This is the surgical alternative to #23758, which reverts the entire
#23533 (~4000 lines). The model code, tool/reasoning parsers, and tuned
MoE configs from #23533 are not part of the bug.

Also re-enables `test_nvidia_nemotron_3_nano` (the stop-gap disable was
added in #23720 when this IMA started showing up).

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

After sgl-project#23019 moved the MoE config loader and the configs/ tree from
`fused_moe_triton/` to `moe_runner/triton_utils/`, two later PRs
unknowingly added 33 tuned-config JSONs to the OLD path:

- sgl-project#22791 (LFM2)        — 24 files (E=32/64, H100/B200/MI325X)
- sgl-project#23533 (Hy3 preview) —  9 files (E=192,N=192 incl. _down,
                                    H20/H20-3e/B200)

The runtime loader anchors its search via
os.path.dirname(os.path.realpath(__file__)) of the loader file
(now in moe_runner/triton_utils/), so configs in the old
directory were never read — runtime fell back to
get_default_config().

The configs themselves were properly tuned and benchmarked at
submission time via the in-process override_config() path used
by the tuning script — that is why the PR authors observed real
speedup. The bug is purely a wrong filesystem location.

Root cause: the tuning README still pointed contributors to the
old path. This PR moves the misplaced configs into the
runtime-loaded location and fixes the README.

Changes:
  * R100 git-mv 33 JSONs into moe_runner/triton_utils/configs/{triton_3_5_1,triton_3_6_0}/
  * Update benchmark/kernels/fused_moe_triton/README.md path

No content changes. No code changes.

References: sgl-project#23019 sgl-project#22791 sgl-project#23533

After sgl-project#23019 moved the MoE config loader and the configs/ tree from
`fused_moe_triton/` to `moe_runner/triton_utils/`, two later PRs
unknowingly added 33 tuned-config JSONs to the OLD path:

- sgl-project#22791 (LFM2)        — 24 files (E=32/64, H100/B200/MI325X)
- sgl-project#23533 (Hy3 preview) —  9 files (E=192,N=192 incl. _down,
                                    H20/H20-3e/B200)

The runtime loader anchors its search via
os.path.dirname(os.path.realpath(__file__)) of the loader file
(now in moe_runner/triton_utils/), so configs in the old
directory were never read — runtime fell back to
get_default_config().

The configs themselves were properly tuned and benchmarked at
submission time via the in-process override_config() path used
by the tuning script — that is why the PR authors observed real
speedup. The bug is purely a wrong filesystem location.

Root cause: the tuning README still pointed contributors to the
old path. This PR moves the misplaced configs into the
runtime-loaded location and fixes the README.

Changes:
  * R100 git-mv 33 JSONs into moe_runner/triton_utils/configs/{triton_3_5_1,triton_3_6_0}/
  * Update benchmark/kernels/fused_moe_triton/README.md path

No content changes. No code changes.

References: sgl-project#23019 sgl-project#22791 sgl-project#23533