feat: add gated synchronous LoadCollection warmup

[sunby]

Summary

Add an opt-in LoadCollection parameter, load_params["warmup"]="sync", with Go SDK WithSyncWarmup() and REST V2 forwarding. Omission keeps existing behavior. Explicit sync rejects refresh, external collections, and in-place enablement on an ordinarily loaded/loading collection; release first.

Persist a separate sync-warmup policy and durable epoch until ReleaseCollection, carry them through load configuration, WAL, query views and segment load information, and qualify physical/transform-loaded fast paths by the successful load epoch. Add stale-view/config-version guards and target-based vchannel × replica progress accounting.

Wire the native force flag into selected field/index/text/JSON-stat/system-field warmup paths, including both outer and inner index warmup settings. This does not change global warmup, lazy or mmap settings, expand selected load fields, or provide a cache-residency guarantee.

Context propagation is unchanged from the PR base. Reuse the existing JsonKeyStats::Load, column-group/helper signatures and cache-slot call sites; force sync uses the existing synchronous execution mechanism. Task-level checks that reject canceled load/update completions remain. Extending native context propagation or prompt I/O cancellation is a separate follow-up, not part of this feature.

Rollout boundary

Draft: native and end-to-end acceptance are still pending. queryCoord.enableLoadCollectionSyncWarmup remains false by default. Admission and placement check QueryNode/native capabilities. A full compatible Proxy/WAL rollout is a manual prerequisite; the QueryNode capability check does not detect every older upstream component. Do not enable the gate based on these unit tests alone.

Scope and references

• Base: sunby/milvus:codex/load-1m-segments-pr-stack-rebased-qv-work. This PR contains only the warmup implementation, not the already-merged query-view resolution work.
• Related issue: [Enhancement]: support user specified warmup milvus-io/milvus#47371 (fine-grained warmup controls; this is a separate load-time override).
• Design: https://github.com/sunby/milvus/blob/b2b2c343926f37e3067f547889b31f30a04216fb/docs/design-docs/design_docs/20260907-load-collection-force-sync-warmup.md
• Companion PyMilvus change: feat: add synchronous LoadCollection warmup option milvus-io/pymilvus#3786

Verification

Checks rerun for the context-scope reduction (b2b2c34392):

• Source diff audit: JsonKeyStats.cpp/.h, ChunkedColumnGroup.h and ChunkedSegmentSealedImpl.h now exactly match the PR base. ChunkedSegmentSealedImpl.cpp keeps only three force-policy hunks; no newly added native context forwarding remains.
• ChunkedSegmentSealedImpl.cpp and JsonKeyStats.cpp passed -fsyntax-only using the current LLVM 21 compile commands, with PCH disabled and OpenMP retained. This is a syntax/signature check, not a full native build or test run.
• run_clang_format.sh passed with clang-format 15.0.7 in an isolated worktree. Changed-line formatting was also checked. Unrelated baseline formatting is excluded.
• make lint-fix ran gofumpt/gci, then failed in root-module type checking: internal/metastore/kv/querycoord/kv_catalog_test.go:372: undefined: mocks. The same reference exists in the base commit and this file is unchanged by the feature. The target did not reach its pkg/client lint stages; full-repository lint is not green.
• git diff --check passed.

Go checks from the previous submission (not rerun for this native-only scope adjustment):

• Full package tests passed for internal/views/coord/loadmgr, internal/views/coord/balancer, internal/views/coord/coordview and internal/views/qviews/....
• Go SDK TestLoadCollectionSyncWarmupOption passed.
• Targeted QueryCoord sync-policy, fail-closed capability and complete-target progress tests passed; the session capability round-trip and isolated default-off gate tests also passed.
• All Go tests above used -tags dynamic,test -gcflags="all=-N -l" -count=1.

Earlier implementation checks and their exact scope are recorded in design appendix B: targeted QueryCoord policy/capability/progress, session capability, default-off configuration, proto and selected native compilation checks. They are not full native/E2E acceptance.

Known gaps / adversarial review

• QN/Proxy/REST test execution is blocked by stale installed native artifacts and missing current native symbols. Their new tests are written, not passed.
• The earlier full milvus_core build attempt encountered Knowhere/LLVM 18/libc++ std::atomic_ref incompatibility, and the chunked-segment build encountered an OpenMP structured-binding capture limitation. The current LLVM 21 syntax checks above pass with OpenMP retained; a full native build/test run was not repeated in this scope-reduction update.
• Full QueryCoord regression is not green: TestServer/TestStop reproduced a repeated-close panic; TestServer/TestUpdateAutoBalanceConfigLoop reproduced an etcd session CAS conflict during setup. Full paramtable tests also encounter local-config default overrides.
• Real Storage V1/V2/V3 I/O barriers, nested Knowhere, text/JSON stats, recovery/migration and concurrent release/reload remain unverified end to end.
• S3 throttling, corruption, OOM, cancellation/timeout, nested-pool/peak-memory behavior and failure-code/retry propagation still need real fault injection. Cancellation tests must verify rejection of late results through the existing cancellation paths; expanding context propagation to promptly interrupt every metadata/file read is outside this feature.
• Full make test-go has not been run for this feature. Keep the gate closed until the acceptance matrix is completed.