Optimize CPU operand materialization with blocked pooled copies

[exAClior]

Context

Profiling the Complex64 TDVP contractions from #53 / #54 found that the remaining right-environment gap is not in faer's complex GEMM. It is in CPU operand materialization.

For abcd,bde->ace and abcde,cef->abdf, OMEinsum must pack the left operand. The current path:

1. default-initializes fresh scratch storage;
2. walks coordinates with a scalar permutation loop;
3. clones the complete packed payload into another fresh Vec;
4. drops both allocations after the contraction.

Tenferro performs one blocked permutation into reusable thread-local storage.

At chi=64, OMEinsum's pre-fix pack measured 137/424 us versus Tenferro's 16/31 us. Prepacking outside the public call removed 87%/97% of the total gaps. OMEinsum also requested two full payload allocations per call: 2.36 MiB for h1-right and 4.72 MiB for h2-right.

Scope

Optimize CPU operand materialization independently of the Complex64 GEMM dispatch work in #54:

• use a blocked, cache-aware strided permutation;
• write directly into the owned packed operand (no second full clone);
• reuse initialized packing storage with a bounded per-thread pool;
• preserve scalar, singleton, padded-stride, arbitrary-permutation, and zero-sized layouts;
• keep the copy path sequential; GEMM threading policy is separate.

Acceptance criteria

• Existing CPU contraction tests pass.
• Add permutation-vs-scalar-reference coverage across representative layouts.
• Add a zero-sized-layout regression test.
• Add packing-buffer reuse coverage.
• make check and cargo package --allow-dirty pass.
• Warm right-layout packing no longer allocates a payload-sized buffer per call.
• Rerun the preserved single-thread comparison with RAYON_NUM_THREADS=1.

Measured target

A local prototype using strided-perm, direct buffer ownership, and a bounded thread-local pool improved OMEinsum by 8-20% in the reverse-order all-size run. Post-fix OMEinsum/Tenferro ratios were 1.00-1.04x across all six right-layout cases, and warm allocation deltas versus prepacked input fell to 800/944 metadata bytes.

This issue deliberately excludes GEMM dispatch and benchmark-harness changes; those remain in #53 / #54.

[exAClior]

Implementation is in draft PR #56, stacked on #54. The PR commit is isolated to Cargo.toml, src/backend/cpu/buffer_pool.rs, and src/backend/cpu/contract.rs; benchmark and profiling changes remain outside it.