[Feature]: Greater CUTLASS Integration #16

What problem does this solve?

• cudaMalloc and cudaFree round-trip passes in energy/sum reduce
• Sequential softmax
• Unoptimized activations

Proposed solution

• Persistent scalar scratch buffer — kills the cudaMalloc/cudaFree round-trip in the energy/sum reductions.
• Softmax kernel parallelized — one block per row instead of one thread, with a caught-and-fixed race in the reduction.
• GELU/Tanh/Sigmoid CUTLASS epilogues.

Alternatives considered

No response

Area

CUDA backend

Checklist

• I checked Roadmap.md and this isn't already tracked there.
• I'm willing to help implement this (optional, just helps with prioritization).