fix(moe): correct MoE grouped-matmul graphs to match hipDNN contract

[SamuelReeder]

Summary

All MoE grouped-matmul graphs in Workloads/headline/moe.tar.gz and
Workloads/microbench/cudnn_bench_moe.tar.gz failed opgraph-level
deserialization against hipDNN. This corrects the graphs to match hipDNN's
MoE grouped-matmul node contract.

Risk assessment

Low. This change updates two DVC-tracked workload archives and their
internal manifests. It does not change hipDNN or provider source code.

Testing summary

• tools/check_deserialize.py --level opgraph (built hipDNN, from_json +
  validate + build_operation_graph): 191/191 graphs pass, fail=0.
• Verified after packaging round-trip (extract from the committed tarballs).
• Confirmed no MoE forward/backward nodes remain with rank-2 tensors,
  non-INT32 routing tensors, or a non-zero top_k outside SCATTER mode.

Testing checklist

• Validate opgraph-level deserialization on both tarballs.
• Validate after packaging round-trip.
• Run the applicability sweep on gfx942 and gfx950.

Technical changes

• Reshaped tensors to hipDNN's required rank-3 convention
  ([1, T, K] / [E, K, N] / [E, 1, 1]).
• Corrected first_token_offset from an E+1-length cumulative-boundary
  array to the E-length per-expert start-offset array hipDNN expects.
• Zeroed top_k on NONE/GATHER mode nodes (only SCATTER uses it).
• Removed a token_ks input incorrectly carried by GATHER-mode dgrad
  nodes (forbidden in that mode).
• Corrected token_index/token_ks tensors to INT32.
• Split graphs with a disconnected router-GEMM component into separate
  *__router.json files (hipDNN requires one connected component per graph).
• Corrected SCATTER-mode output row count to the dispatched-token count.
• Excluded 21 wgrad graphs whose token/doutput row counts mismatch
  (pre- vs post-scatter counts); the current backward node has no
  scatter/token_index support, so this isn't a shape fix. Documented as a
  known gap in the manifest.

[SamuelReeder]

Superseded — the moe fix is now part of #39 (norm + moe on one branch), no need for a separate PR for these DVC-only changes.