Optimize CPU operand materialization with blocked pooled copies

[exAClior]

Summary

• replace the scalar coordinate-carry materializer with strided-perm's blocked copy
• return the packed destination directly instead of cloning the full payload
• reuse initialized packing storage through a bounded thread-local best-fit pool
• guard zero-element layouts before entering the blocked kernel
• add scalar-reference permutation, zero-size, bounded-allocation, and buffer-reuse tests

Closes #55.

Why this is separate

This is stacked on #54 because the profiling was discovered while validating its Complex64 contractions, but the fix is generic CPU operand materialization. It is deliberately a separate commit and PR: no Tenferro dependency, benchmark harness, profiling source, Makefile change, or run artifact is included.

Until #54 lands, GitHub will show its prerequisite commits in this PR's main-branch comparison. The materialization change itself is the single commit cee9fae and touches only:

• Cargo.toml
• src/backend/cpu/buffer_pool.rs
• src/backend/cpu/contract.rs

Evidence

The pre-fix right-layout pack took 137/424 us at chi=64, versus Tenferro's 16/31 us. Prepacking removed 87%/97% of the public-path gap.

With this patch, the reverse-order all-size run improved OMEinsum by 8-20%. Post-fix OMEinsum/Tenferro ratios were 1.00-1.04x across all six right-layout cases. At chi=64:

Case	OMEinsum before	OMEinsum after	Tenferro after
h1-right	1,053.6 us	906.6 us	878.8 us
h2-right	2,145.5 us	1,760.0 us	1,738.9 us

Warm allocation deltas versus prepacked input fell from 2,359,616/4,718,992 bytes to 800/944 metadata bytes.

The first blocked-copy implementation exposed a SIGSEGV for a zero-sized layout. The final code bypasses the external kernel for numel == 0, with a dedicated regression test; the existing zero-dimension integration test also passes.

Verification

• make check
  • 162 unit tests passed; 11 ignored
  • 333 integration tests passed
  • 16 doctests passed; 4 ignored
• cargo package --allow-dirty
• exhaustive small-rank permutation checks against the scalar reference

Depends on #54.

[exAClior]

Isolated stacked diff against the #54 head branch: exAClior/omeinsum-rs@perf/complex-gemm...perf/issue-55-materialization — three files, one commit (cee9fae).

[codecov-commenter]

⚠️ Please install the to ensure uploads and comments are reliably processed by Codecov.

Codecov Report

✅ All modified and coverable lines are covered by tests.

📢 Thoughts on this report? Let us know!

[exAClior]

Scope and memory-retention note: This optimization affects CPU operand materialization across all scalar types and algebra paths, not only large Complex64/TDVP workloads. The blocked copy is expected to help large permutations, but the crossover for tiny permutations has not yet been benchmarked. The thread-local cache may retain up to two 64 MiB buffers per scalar type per thread—a theoretical 128 MiB/type/thread high-water mark. The current contraction path generally needs only one packing buffer at a time, so please advise whether the pool should retain only one buffer or use a lower cap before merge.