Spike: investigate SBUF residency for small-op dispatch overhead

[scttfrdmn]

Parent: #40. Research spike before implementation.

v0.7.0/v0.8.0 benchmarks show small ops (mask 64×32, GEMM 128, ComplexLinear 128→256) are 30-100× slower on NKI than CPU because of per-call XLA dispatch + tensor staging overhead.

Investigate:

• Can intermediate tensors stay in SBUF across multiple kernel invocations?
• What's the actual overhead breakdown (XLA trace time, HBM round-trip, kernel launch)?
• Would batching short sequences of small ops into a single fused kernel help?

Deliverable: a decision memo (pick approach, or conclude the overhead is irreducible without a larger architectural change).