…#24179)

### Background

PR #24153 introduced an `OldColCapture` mechanism in DEDUP JOIN to merge REPLACE INTO's two main-table scans into one. Each worker recorded probe-side old-column values into per-worker `capturedVecs` keyed by build bucket, intended to be emitted alongside the build row at finalize.

That introduced a latent bug: parallel probe workers kept `capturedVecs` and `captured` as container-local state, but the `finalize()` multi-worker protocol only merged the `matched` bitmap through `ap.Channel`. Non-merger workers' captured values were silently dropped, so rows matched by those workers would emit NULL (or stale) in the placeholder slot.

Commit `7cc544d2d` (`fix(replace): disable parallel probe for DEDUP JOIN with OldColCapture`) worked around this by forcing `Mcpu = 1` on every probe scope whenever `OldColCapture` was active (`pkg/sql/compile/compile.go`). Shuffle + capture additionally panics NYI. The REPLACE INTO merged-scan path has thus been running **without any probe-side parallelism** since.

### This PR

Restore parallelism by upgrading the finalize merge protocol.

1. New `WorkerJoinMsg` struct in `pkg/sql/colexec/dedupjoin/types.go` carries `{matched, captured, capturedVecs}`. `DedupJoin.Channel` becomes `chan *WorkerJoinMsg`.
2. Non-merger workers relinquish ownership of their capture state to the merger when sending; the merger:
- Or's in the matched bitmap (existing behavior).
- Walks the worker's `captured` bitmap and, for each bucket not already present in its own `captured`, copies per-column values from the worker's `capturedVecs` into its own. First-wins semantics across workers — any one captured value for a bucket is semantically equivalent since HashOnUnique gives a 1:1 bucket↔build-row mapping.
- Frees the worker's `capturedVecs` after merging (ownership was transferred).
3. Remove the `Mcpu = 1` forcing in `compile.go` so broadcast DedupJoin with OldColCapture runs parallel again. The shuffle + capture NYI panic stays — cross-CN pipeline channel semantics for `capturedVecs` are out of scope here and will be addressed separately.
4. The defensive capture-field copy in `dupOperator` (also from `7cc544d2d`) is retained as it's a prerequisite for parallel clones.

### Tests

Added unit tests in `pkg/sql/colexec/dedupjoin/join_test.go`:

- `TestMergeCaptured_DisjointBuckets` — parallel workers capture different buckets; after merge, merger owns the union.
- `TestMergeCaptured_FirstWinsOnConflict` — when both workers captured the same bucket, merger keeps its own value.
- `TestMergeCaptured_EmptyWorkerMsg` — worker with empty capture doesn't corrupt merger state.
- `TestWorkerJoinMsg_ChannelRoundTrip` — full channel send/receive + merge + free cycle without leaks.
- `TestReceiveWorkerMsg_ContextCancel` / `_ChannelClose` — receive helper respects context cancellation and closed channel.

Existing end-to-end `TestDedupJoinCapture{,PartialMatch,Reset}` unchanged, all pass. `go test -race` clean. `make static-check` clean.

### Benchmark

YCSB `workload-replace` (100K ops, updateproportion=1, zipfian, threads=16, single CN, one iteration):

| metric                   | baseline (main) | this PR   | delta    |
|--------------------------|----------------:|----------:|---------:|
| RUN Throughput (ops/sec) | 1384.64         | 1561.67   | +12.8%   |
| UPDATE avg latency (us)  | 11255.67        | 9978.82   | -11.3%   |
| UPDATE p95 (us)          | 22527           | 18223     | -19.1%   |
| UPDATE p99 (us)          | 53471           | 40895     | -23.5%   |
| UPDATE max (us)          | 593407          | 260095    | -56.2%   |

Tail latency improvements (p99 -23%, max -56%) match the expected effect of restoring parallel probe: reduced queueing at the DEDUP JOIN stage under the high-conflict zipfian workload.

### Follow-ups (separate issues / PRs)

- Enable shuffle + capture (needs cross-CN `Channel`/`WorkerJoinMsg` transport).
- Re-evaluate PR #24163 (REPLACE static scan pushdown + IF-guard index rewrite) under the restored parallel baseline.

Approved by: @aunjgr