operations: support background execution of long-running operations

Some operations (e.g., `backup-restore`, `add-index`, `manual-compaction`) run for hours and block a regular worker for the entire duration, or time out before completing. These operations reduce runner throughput by occupying limited worker slots.

**Proposal:** Add a `LongRunning bool` field to `OperationSpec`. Operations marked as long-running are executed by a separate pool of background workers with a configurable concurrency limit (e.g., `--long-running-operation-parallelism`). The operation implementation remains unchanged — it runs the same `Run()` and returns cleanup as usual. The only difference is scheduling: long-running operations don't consume regular worker slots.

This keeps the regular worker pool available for short, frequent operations while giving long-running operations dedicated capacity.

**Key files:** `pkg/cmd/roachtest/run_operation.go`, `pkg/cmd/roachtest/registry/operation_spec.go`

Jira issue: CRDB-61915