Add: benchmark scene tests for dispatch and graph topologies

[chenshengxin2026]

Summary

• Add 6 new scene tests for tensormap_and_ringbuffer runtime covering dispatch overhead and graph topology benchmarks:
  • Dispatch-independent: measures dispatch overhead as task count scales with independent scheduling
  • Dispatch-serial: measures serial dispatch throughput
  • Graph-chain_N: chain topology (A → B → C → ...)
  • Graph-fanin_N: fan-in topology (multiple sources → single sink)
  • Graph-fanout_N: fan-out topology (single source → multiple sinks)
  • Graph-diamond: fork-join topology (A → {B_0..B_W} → D)
• Restructure benchmark_bgemm test cases with descriptive parameter sweep naming (Tile, Batch, K, Loop sweeps) replacing generic Case0-4
• Register all new scene tests in benchmark_rounds.sh

Test plan

• Run Dispatch-independent and Dispatch-serial on device to verify dispatch overhead measurement
• Run Graph-chain_N, Graph-fanin_N, Graph-fanout_N, Graph-diamond on device to verify graph topology tests
• Run benchmark_bgemm with new case names to verify restructured parameter sweeps
• Run full benchmark rounds to confirm all new entries execute correctly

[gemini-code-assist]

Code Review

This pull request introduces a comprehensive suite of performance and topology tests for the tensormap_and_ringbuffer runtime, including benchmarks for independent and serial dispatch, as well as various graph structures such as linear chains, diamonds, fan-ins, and fan-outs. Additionally, it expands the benchmark_bgemm test cases to include sweeps for tile size, batch groups, and dimensions, and integrates these new tests into the benchmark tools. The review feedback correctly identifies several discrepancies between the docstrings and the actual implemented test cases in the golden scripts, providing specific suggestions to align the documentation with the code.