feat(comm): add copy-engine selected-record exchange with bulk layer packets

[FujitsuPolycom]

Summary

Adds a generic copy-engine transport for sparse fixed-width records.

• exchange() moves one compact record plane.
• exchange_layers() packs three planes with one destination map, performs one
  rank-rotated peer DMA sequence and one arrival barrier, then unpacks into
  caller-owned outputs.
• Empty, single-plane, and layered calls share a graph-safe deferred-release
  protocol and can alternate on one stream.
• Pool-scaled offsets use 64-bit arithmetic.

The public API is sparkinfer.comm.pcie.SelectedRecordCopyExchange. This PR is
stacked on the direct selected-record transport so both implementations remain
independently reviewable.

Sparkinfer rebase

The branch is based on local-inference-lab/sparkinfer@ec2cd4d; all runtime,
JIT, packaging, and test paths use the new sparkinfer.comm.pcie namespace.

Validation

• focused CPU/source suite: 33 passed, 2 GPU-gated skips
• ai01 exact-image DCP4/432-byte GPU gate: 1 passed
• the DCP8/656-byte case is retained and skipped on the four-GPU host
• eager exchange, CUDA graph replay, overflow, mixed replay, and >2 GiB offsets passed
• packaged-wheel import, Ruff, compileall, and vLLM policy tests passed
• rebased v19 integration image: 120 passed, 18 skips in CKV/indexer policy tests

Performance context

The dependent vLLM integration groups GLM S1/S2/S3 selected-record prefetches
into one exchange. Against the same-image per-layer CE control:

• 64K+ comparable decode geometric mean: +8.50%
• coding median: 99.63 -> 103.79 tok/s
• reported KV capacity: unchanged at 302,047 tokens

Operational contract

• every rank calls exchanges in identical order
• instances are stream-affine
• eager warmup precedes CUDA graph capture
• callers fall back to single-plane exchange above layered capacity

AI-assisted implementation under human direction; the diff and validation
results were reviewed before publication.

Design document

GLM-5.2 Sparse CKV under Decode Context Parallelism is the canonical architecture, transport-contract, validation, and upstream-decomposition reference for this PR stack.