Clean up Oink benchmarks and document SM103 results

[Laurawly]

Summary

This PR cleans up the Oink benchmark/docs surface and updates the repo for current SM10x / GB300 usage.
Changes include:

• Make the top-level Oink README more concise and current.
• Update benchmark docs for SM103 / GB300 setup and measured roofline context.
• Add DSv4 norm benchmark coverage for RMSNorm, RMSNorm backward, fused-add RMSNorm, and LayerNorm.
• Add fused-add RMSNorm DSv3/DSv4 speedup tables vs Quack to benchmarks/README.md.
• Generate and include a current SM103 BF16 benchmark SVG.
• Centralize CuTeDSL cache-dir setup in _cutedsl_cache.py.
• Update package/docs wording from SM100-only to Blackwell SM10x / SM100-SM103 where appropriate.

Validation

Ran on GB300 / SM103 with CUTE_DSL_ARCH=sm_103a:

git diff --check
PYTHONNOUSERSITE=1 CUTE_DSL_ARCH=sm_103a python -m compileall -q src benchmarks tests
PYTHONNOUSERSITE=1 CUTE_DSL_ARCH=sm_103a python -m pytest tests -q

Result:

15 passed