kvserver/rangefeed: RangeFeedProcessors gauge can go negative when Start fails after scheduler registration

**Describe the problem**

`ScheduledProcessor.Start` increments `RangeFeedProcessors` only after all its fallible steps ([scheduled_processor.go#L152](https://github.com/cockroachdb/cockroach/blob/a7e117882b4356096b3a8ea26046baa7e5d53fce/pkg/kv/kvserver/rangefeed/scheduled_processor.go#L152)), but the error paths in between (rtsIter construction failure, `RunAsyncTask` failure) call `p.scheduler.StopProcessor()` after the scheduler registration has already succeeded. The eventual `processStop` decrements the gauge unconditionally ([scheduled_processor.go#L254](https://github.com/cockroachdb/cockroach/blob/a7e117882b4356096b3a8ea26046baa7e5d53fce/pkg/kv/kvserver/rangefeed/scheduled_processor.go#L254)), so each such failure drives the gauge one below reality, permanently.

Metric-only impact.

**Suggested fix**

Move the `Inc` before the fallible steps, or make `processStop` decrement only when `Start` completed.

**Additional context**

Found during an agent-assisted correctness audit of the rangefeed subsystem. Code links are pinned to master @ a7e1178.

Jira issue: CRDB-65656

Start in pkg/kv/kvserver/rangefeed/scheduled_processor.go, reading ScheduledProcessor.Start around line 152 and processStop around line 254. Trace the rtsIter construction and RunAsyncTask failure paths after scheduler registration, then verify that each failed start leaves RangeFeedProcessors accurate and unable to go negative.

Written by the indexing model from the issue text.