Copy allocated arrays through backend-owned transfers

[PraneethMerugu]

Summary

• copy allocated arrays through the owning backend instead of invoking Base's
  generic destination copy on device storage
• preserve the unchanged synchronous Base copy for CPU storage
• materialize unsupported ordinary host views before crossing the backend
  transfer boundary while retaining native device-to-device transfers
• exercise independent ownership for reversed CPU views, strided structured
  host views on Metal, and native MtlArray sources
• document the qualified CPU/Metal view-allocation contract

Why this is required

CorePotts' structured maintained-quantity rebuild supplies ordinary array views
to LocalMath's existing Allocate(source) boundary. The former generic
copyto! implementation attempted scalar indexing when that host view targeted
Metal storage. This change corrects the sole cold allocation copier; it adds no
second executor, storage authority, or backend registry.

Validation

• full ordinary LocalMath suite: 1,735/1,735
• focused storage, package-quality, and runner-inventory cohort: passed
• complete actual-Metal owner suite with scalar indexing disabled: passed
• focused Metal authoring and transfer fixture: 37/37
• downstream CorePotts structured maintained-sum Metal fixture: 24/24
• strict documentation build: passed
• independent review: no blocking finding
• git diff --check: passed

CPU and Metal are the qualified paths. Dispatch applicability is not presented
as a blanket capability guarantee for CUDA, ROCm, custom backends, or arbitrary
provider array wrappers; those combinations retain their own later conformance
work.