chore(nccl): migrate internal/nccl to purego/dlopen

[dndungu]

Closes #78.

Replaces the CGo binding in internal/nccl with a runtime dlopen of libnccl.so.2, mirroring the pattern in internal/cublas/cublas_purego.go. The package now compiles on every platform without -tags cuda; non-linux GOOS returns a clean "not supported" error rather than panicking.

Summary

• New internal/nccl/nccl_purego.go resolves ncclGetUniqueId, ncclCommInitRank, ncclCommDestroy, ncclCommGetAsyncError, ncclAllReduce, ncclBroadcast, ncclGroupStart, ncclGroupEnd, and ncclGetErrorString via cuda.DlopenPath/cuda.Dlsym/cuda.Ccall. ABI constants for ncclResult, ncclDataType, ncclRedOp, and NCCL_UNIQUE_ID_BYTES = 128 are hardcoded against the stable NCCL 2.x ABI.
• Legacy CGo file moved to internal/nccl/nccl_cgo.go and gated behind //go:build cuda && cgo && nccl_cgo (OFF by default; opt-in fallback only).
• nccl_test.go no longer requires -tags cuda. A requireNccl(t) helper skips when libnccl.so.2 is not dlopen-able. Two new tests (TestConstants, TestUniqueIDFromBytesRoundTripNoLib) exercise the pure-Go marshaling and ABI constants on every platform.
• ADR-002 documents the rationale and the AAPCS64 hidden-pointer trick that lets us pass the 128-byte ncclUniqueId by value through the shared cuda.Ccall trampoline.
• CI vet exclude list adds /internal/nccl$ alongside the other GPU runtime binding packages.

AArch64 ABI note

ncclCommInitRank takes ncclUniqueId by value (128 bytes). Per AAPCS64 rule B.4, composites larger than 16 bytes are passed by hidden caller-allocated pointer, so passing uintptr(unsafe.Pointer(&uid.id[0])) is the correct calling convention on linux/arm64 (the only NCCL platform ztensor targets today). If we ever need linux/amd64 NCCL, the SysV ABI passes large aggregates on the stack and we will need either an assembly trampoline or the nccl_cgo fallback.

Verification on DGX (linux/arm64, real libnccl.so.2)

=== RUN   TestConstants
--- PASS: TestConstants (0.00s)
=== RUN   TestUniqueIDFromBytesRoundTripNoLib
--- PASS: TestUniqueIDFromBytesRoundTripNoLib (0.00s)
=== RUN   TestGetUniqueID
--- PASS: TestGetUniqueID (0.01s)
=== RUN   TestUniqueIDRoundTrip
--- PASS: TestUniqueIDRoundTrip (0.00s)
=== RUN   TestUniqueIDFromBytesInvalidLength
--- PASS: TestUniqueIDFromBytesInvalidLength (0.00s)
=== RUN   TestSingleRankInitDestroy
--- PASS: TestSingleRankInitDestroy (1.15s)
=== RUN   TestSingleRankAllReduce
--- PASS: TestSingleRankAllReduce (0.52s)
=== RUN   TestGroupStartEnd
--- PASS: TestGroupStartEnd (0.00s)
PASS
ok  	github.com/zerfoo/ztensor/internal/nccl	1.816s

(Two-GPU tests skip on the single-GPU DGX Spark host as expected.)

Test plan

• go build ./... on darwin/arm64 (no tags) — clean
• go test ./internal/nccl/... on darwin/arm64 — pure tests pass, NCCL tests skip
• go build ./... on linux/arm64 DGX — clean
• go test ./internal/nccl/... on linux/arm64 DGX — all NCCL paths exercised against real libnccl.so.2
• CI green

Follow-up

• The duplicate internal/nccl copy inside the zerfoo repository is not touched by this PR. It should be migrated separately.