feat(dataobj): Schedule LogMerge via per-tenant IndexMerge/LogMerge flip-flop

[trevorwhitney]

What this PR does

Changes how dataobj log compaction (LogMerge) is scheduled. Previously LogMerge only ran once a (tenant, window) had converged to a single index (len(entries) == 1 routing in the old poll loop). Under continuous ingestion a window routinely holds more than one index, so it never converged and its logs were never compacted.

This replaces the cron/poll loop with a per-tenant IndexMerge ↔ LogMerge flip-flop:

• coordinator.Run spawns one worker goroutine per configured tenant (-dataobj.compaction.tenants) and drains them via a WaitGroup on shutdown. There is no poll loop.
• Each worker alternates phases: IndexMerge (unchanged compactTenant) then LogMerge. On success it flips; on error it re-arms the same phase.
• LogMerge no longer waits for convergence — it compacts the log objects behind every index in the window, one index at a time, reusing the existing compactTenantLogs per index. This is the core fix.

Why

LogMerge effectively never triggered in environments that don't converge to a single index. Decoupling it from convergence lets log compaction make progress regardless of index count.

Scope / notes

Targeted at getting LogMerge running in a dev environment for a single configured tenant. Deliberately deferred (documented for revisit before broader/production use):

• Never-sleep loop: maximizes compaction utilization; the intended backpressure is scheduler-queue depth (not built here). An idle/converged tenant will hot-loop ToC reads — acceptable for a single active dev tenant.
• Multi-tenant ToC-swap serialization (single shared per-window ToC object) — a process-global swap mutex is deferred; safe with one worker.
• PollingInterval is retained as a config knob but no longer read by the scheduler (reserved for a future inter-phase throttle).
• Metrics migration is minimal: existing counters are reused; "cycle" metrics now count per-tenant phase iterations (compacted|converged|failed).

Config

• New: -dataobj.compaction.tenants (comma-separated) — the tenants to compact; one worker each. Static (restart to change); empty = no workers.

Testing

• Unit tests for both phases (per-index LogMerge fan-out incl. delta-counter assertions, phaseOutcome mapping, cancellation short-circuit), plus lifecycle tests for Run/runTenantLoop (WaitGroup drain on cancel, error-keeps-phase / success-flips, tenant dedupe).
• go test ./pkg/engine/... ./pkg/dataobj/... green; go vet and gofmt clean.

Operational acceptance (dev)

With -dataobj.compaction.enabled=true -dataobj.compaction.tenants=<tenant>, loki_dataobj_compaction_tenant_log_cycles_total{outcome="compacted"} should increment on windows with >1 index (which previously never happened), and ..._indexes_added_total / ..._tasks_total advance during log-merge phases.