Settle validation through provider completion copies

[PraneethMerugu]

Summary

• narrow validation settlement to one program-level device/host status pair
• use a nonempty blocking validation copy as the provider-scope completion operation, with explicit synchronize fallbacks for empty or aliased storage
• cache provider failures and release exact receipt leases on wait and waitall
• reuse the same validation-transfer owner for fixed-relation admission
• preserve cumulative provider ordering, deterministic argument-order failures, public receipt semantics, and one CPU/Metal KernelAbstractions path

Compiler and allocation evidence

• pending wait: 208 to 176 bytes; combined and grouped waits improve by 32 bytes at every measured arity 0/1/2/4/17
• execute unchanged at 416 bytes; cached wait remains 0 bytes
• broad wait typed IR: 131/76/36 to 130/75/35 statements/calls/Any
• one MethodInstance each for receipt settlement, prepared-scope settlement, and lane-tail settlement; no dependency-arity growth
• concrete JET clean; AllocCheck reports only the known IdSet insertion ccall

Validation

• focused CPU: receipts 79/79, structural 80/80, storage authoring 80, authoring 98/98, SourceOrder 80/80, ordered-fold control 26/26
• full CPU: 1,943/1,943 across 47 testsets
• focused real Metal: 137/137 including copy-surfaced device failure, fixed relation, and SourceOrder
• full real Metal: 592/592 across 27 testsets
• independent architecture review clean

The change adds no cache, alternate executor, backend-specific scientific path, or second status/scheduling authority.