feat(vchordrq): add pluggable exact MaxSim reranking

[HuXinjing]

Summary

This is the second implementation slice proposed in Discussion #476.

It introduces a storage-independent exact MaxSim reranking boundary and a CPU reference backend. The reference tensor source reads the indexed expression from the PostgreSQL heap under the active scan snapshot; it does not introduce an external registry, network transport, GPU dependency, or quantized format.

User-visible behavior

• Existing behavior remains the default through vchordrq.maxsim_backend = 'coarse_only'.
• Exact reranking is opt-in with vchordrq.maxsim_backend = 'cpu_exact'.
• Exact mode requires an explicit positive vchordrq.maxsim_candidate_limit, bounded at 65,536.
• Candidate generation remains the existing vchordrq MaxSim path. Only the bounded candidate set is rescored using the full heap tensor.

Architecture boundary

• CandidateTensorSource separates full-tensor retrieval from scoring.
• ExactMaxsimBackend separates exact scoring from candidate generation and storage.
• HeapTensorSource is the only source in this PR.
• CpuExactMaxsimBackend is the only backend in this PR.

External tensor-source registration and any optional GPU process remain separate follow-up proposals, contingent on maintainer feedback.

Correctness and safety

• Uses the existing HeapFetcher, active PostgreSQL snapshot, and indexed-expression evaluation.
• Rechecks the active scan qual before materializing a candidate tensor.
• Rejects null, mismatched-kind, and mismatched-dimension tensors.
• Does not add background workers, sockets, persistent state, migrations, or build dependencies.

Validation

• cargo fmt --all -- --check
• cargo test --locked --workspace --exclude vchord --no-fail-fast
• PG16 release extension build and generated SQL schema
• PG16 root-extension unit test for exact result ordering
• Added tests/vchordrq/maxsim_cpu_exact.slt

Local validation gap: the isolated PG16 build image contains development headers but no initdb, so the new SQLLogicTest was not executed against a locally started PostgreSQL server. It is expected to run in the repository CI PostgreSQL matrix.

Relationship to other work

• Discussion: RFC: optional exact late-interaction reranking and GPU scoring backends #476
• Planner statistics/cost model: feat(vchordrq): estimate MaxSim planner cost #477

This branch is independently based on current main; it does not depend on #477.

[github-actions]

All contributors have signed the CLA ✍️ ✅
_{Posted by the CLA Assistant Lite bot.}

[HuXinjing]

I have read the CLA Document and I hereby sign the CLA