Add compiled CopyPlan for prepared strided copies (stage 1 of #139)

[Ryo-wtnb11]

Add compiled CopyPlan for prepared strided copies

Summary

Stage 1 of #139: a prepared, reusable copy plan over the raw borrowed-layout kernels from #140. CopyPlan::compile runs the layout planning once (validation + fused-pair layout); execute / execute_scale / execute_conj replay it against RawStridedRef / RawStridedMut with zero heap allocations for rank ≤ 8.

let plan = CopyPlan::compile(dims, dst_strides, src_strides)?; // plan once
plan.execute(dst_raw, src_raw)?;                               // zero-alloc, repeatable
plan.execute_scale(dst_raw, src_raw, alpha)?;

Design

• Layering: execute is exactly "Add allocation-free raw copy/axpy kernels #140 minus the per-call planning" — it reuses the same pub(crate) fused-pair machinery (fuse_pair_layout / apply_fused_pair), so there is no second copy-kernel implementation. Rank > 8 falls back to the view kernels (the explicit-fallback rule; this is the only allocating path).
• Validation split: compile proves layout compatibility (rank agreement, extent-product overflow, destination injectivity via the existing is_injective_layout); execute* re-checks only the per-call facts (the supplied views must carry exactly the compiled layout — PlanLayoutMismatch otherwise). Buffer bounds are already proven by RawStridedRef/Mut construction.
• Public surface kept minimal: only CopyPlan + its methods and the new StridedError::PlanLayoutMismatch variant become public; all planning internals stay pub(crate).

MapPlan / ReducePlan (stages 2–3 of #139) are deliberately not included; this PR proves the API shape on the smallest reviewable slice.

Tests

• Differential plan-vs-direct (byte-identical against the Add allocation-free raw copy/axpy kernels #140 raw kernels): rank 0/1/2/4/8, zero-size dims, identity layout (asserts single fused axis), transposed and negative-stride layouts, rank-9 fallback, f32/f64/c32/c64, scale and conj semantics.
• Compile rejections (length mismatch, extent overflow, non-injective destination; broadcast source allowed) and execute-time layout-drift rejection (dims / dst strides / src strides).
• Zero-allocation contract: a counting global allocator asserts execute / execute_scale / execute_conj perform no heap allocations at rank 8, on both the default and parallel features.

cargo test --workspace green (incl. --features parallel), cargo fmt --check clean, cargo doc --no-deps warning-free, per-file coverage ≥ 80% (copy_plan.rs 99.3%).