chore(llm): update Burn and Cube accelerator stack

[ppodolsky]

Summary

• update the direct dependency bases to Burn v0.22.0-pre.1, CubeCL v0.11.0-pre.1, and CubeK v0.3.0-pre.1
• rebase the required accelerator changes on each current upstream main
• pin the three consolidated upstream PR heads until they merge:
  • feat(cuda): add resilient native and grouped GEMM dispatch tracel-ai/cubecl#1440
  • feat(attention): emit forward LSE and harden matmul tuning tracel-ai/cubek#428
  • feat(cubecl): add native GEMM autotune candidate tracel-ai/burn#5190
• include the CubeK matmul autotune-key fix that separates 16-byte-aligned and unaligned stride layouts within an anchored shape bucket

The legality split fixes a reproducible A100 failure where an async-copy matmul selected for an aligned tensor was cached and reused for an incompatible unaligned tensor. The later Burn fusion-ordering panic was a consequence of that first failed launch.

All Cargo files submitted to the upstream dependency PRs reference only tracel-ai/* repositories. This repository's root patch table temporarily resolves the fork PR heads and can be removed as each upstream PR merges.

Validation

• cargo check -p hermes-train --features cuda
• CubeCL runtime and CUDA-example checks
• CubeK matmul-key regression tests plus attention checks
• Burn CUDA check plus accelerated-GEMM contract tests
• A100 exact-head smoke on the 299,929,088-parameter MoE model: five optimizer steps completed, loss 10.923190 → 10.607383, final throughput 41,997 tokens/s
• two order-reversed 130-step A/B pairs showed that explicitly flushing CubeCL's client dispatch queue regressed throughput by 0.254% (42,104.8 → 41,998.0 tokens/s), so that candidate and its API were removed

No model architecture, training data, curriculum, or checkpoint format changes are included.