kvclient: request state abandoned when a panic unwinds through DistSender.Send and Streamer.Close

**Summary:**

Two paths in `pkg/kv/kvclient` leave state behind when a panic unwinds through them. Go runs deferred functions during unwinding, so the gap here is cleanup written as plain statements. Part of #174944.

**Findings:**

- `kvcoord.DistSender.Send`: `responseChs` — on the panic path the drain is deliberately skipped, immediately below a comment explaining that draining is required because the `client.Sender()` contract forbids holding onto any part of a request after `Send` returns. Per-range goroutines therefore keep reading and mutating `ba` / `ba.Requests` after the caller has moved on and may have recycled or re-accounted those buffers.
- `kvstreamer.Streamer.Close`: `metrics.OperatorsCount` — the gauge decrement and `*s = Streamer{}` are plain statements after `results.close` and `truncationHelper.Release`, so a panic in either leaves the node-level gauge permanently inflated. Observability only; no job-relevant gauge feeds a control decision.

**Next Steps:**

- [ ] `DistSender.Send`: on the panic path, either drain with a bounded timeout or deep-copy what the async paths reference
- [ ] `Streamer.Close`: defer the gauge decrement

Epic: none

Jira issue: CRDB-68117

Start in pkg/kv/kvclient by reading kvcoord.DistSender.Send, especially the panic-path handling around responseChs, and kvstreamer.Streamer.Close around results.close, truncationHelper.Release, and metrics.OperatorsCount. Determine whether bounded draining or deep-copying safely satisfies the Sender contract, then verify that panic unwinding leaves no request state behind and the gauge is decremented.

Written by the indexing model from the issue text.