feat(tasks): add curated Campaign 20 kernel benchmarks

[irvineoy]

Summary

This PR rebuilds the independently mergeable subset of #44 on top of the
current main, instead of carrying forward the original 1,780-file change.

• adds four self-contained Triton tasks:
  • gemm_a16_w16_kernel
  • gemm_a8w8_blockscale_kernel
  • topk_forward
  • write_req_to_token_pool_triton
• adds two self-contained HIP tasks:
  • paged_attention_large
  • wvsplitk
• gives every task a real compile/JIT smoke check, correctness check, and
  canonical performance timing path
• removes runtime dependencies on installed AITER, vLLM, SGLang, and
  triton_kernels packages
• targets MI355X (gfx950) and derives device properties such as CU count at
  runtime
• generalizes opt-in perf-helper materialization to marked task runners
• stores generated workspace_* directories below ./experiments/ and
  git-ignores that directory
• keeps the one-off task-validator config out of the repository

Scope decisions from #44

This intentionally leaves out:

• tasks that already exist on main
• tasks with missing local source or that call production engine operators
• paged_attention_ragged, because its previous form vendored a full
  Composable Kernel tree; it should be a separate pinned repository task or
  vendor only the minimal dependency closure

The resulting task change is under 1 MiB, versus 1,780 files and about 696k
added lines in #44.

Validation completed

Static and framework checks:

• git diff --check
• Python compile/JSON/YAML parsing and task discovery for all added tasks
• config/source/target-symbol contract checks for all six tasks
• quoted local-header dependency closure checks for both HIP tasks
• make check-perf-helpers
• make check-evaluator
• make check-docker-runner
• make check-visualization

Task validator on an AMD Instinct MI355X (gfx950) using
lmsysorg/sglang-rocm:v0.5.14-rocm720-mi35x-20260705:

• all four Triton tasks: PASS, 10/10 checks each
• paged_attention_large: PASS, 10/10 checks
• wvsplitk: PASS, 10/10 checks
• combined final result: 6/6 tasks PASS, with compile, correctness,
  correctness-implementation review, performance, self-contained, and GPU-hang
  checks passing; no warnings or timeouts in the final reports

The validator runs used temporary ignored YAML files, which were deleted after
the runs. Their reports are retained locally below ./experiments/ and are not
part of this PR.

Related to #44.