(feat): add reshape! — zero-allocation array reshaping via pool wrapper cache

[mgyoo86]

Summary

Add reshape!(pool, A, dims...) that reshapes arrays using the pool's N-D wrapper cache. On Julia 1.11+, cross-dimensional reshapes are zero-allocation after warmup via setfield!-based wrapper reuse. The result shares memory with the source array.

Motivation

Base.reshape allocates a new wrapper object on every call (~40-80 bytes). In hot loops that repeatedly reshape pooled arrays (e.g., treating a flat vector as a matrix for BLAS), these wrapper allocations accumulate. reshape! eliminates this by reusing cached wrappers from the pool's existing nd_wrappers infrastructure.

Changes

File	Change
src/acquire.jl	Add _claim_slot! (wrapper-only slot reservation) and _reshape_impl! for AdaptiveArrayPool using setfield!(:ref, :size)
src/convenience.jl	Add public reshape! API with _record_type_touch!, DisabledPool fallbacks, and _impl! delegators
src/macros.jl	Transform reshape! → _reshape_impl! inside @with_pool/@maybe_with_pool; extract eltype(A) for typed-lazy mode
docs/	API docs, quick-start example, and reference table entry
test/test_reshape.jl	427-line test suite (see below)

Design

• Same dimensionality (M == N): setfield!(A, :size, dims) in-place — no pool interaction, effectively free
• Cross dimensionality (M != N): _claim_slot! reserves a pool slot without backing memory, then reuses a cached Array{T,N} wrapper via setfield!(:ref, :size) pointing to A's MemoryRef
• Memory sharing: Reshaped array shares memory with original — mutations visible in both directions
• Validation: DimensionMismatch thrown before slot reservation if prod(dims) != length(A)
• Julia 1.10 fallback: Base _reshape_impl! delegates to Base.reshape