fix: fire setDirty when pager fetch pipeline drains

[onderilkesever]

Problem

Fixes #298.

When SparkRenderer is used in an on-demand rendering architecture (render only when onDirty fires, no continuous loop) with a static camera, paged LOD splats can stop loading partway through. The onDirty chain has a gap: chunk fetches complete asynchronously, outside any render frame, and nothing calls setDirty() to signal that new data is ready.

Root cause

setDirty() is only called from:

1. After sort completes (driveSort)
2. When the LOD snapshot changes (updateInternal)
3. After updateLodInstances when lodDirty = true (driveLod)

All three require a render frame to be in progress. If the camera is static and dirty = false, onBeforeRender never fires → driveLod never runs → consumeLodTreeUpdates is never called → downloaded chunks sit in lodTreeUpdates forever.

Fix

Add an optional onDirty callback to SplatPagerOptions and call it when processFetched() fully empties the fetched queue — i.e. after each chunk finishes decoding and is moved into the LOD pipeline. SparkRenderer passes () => this.setDirty() when constructing the pager.

This fires after every individual chunk lands, giving progressive rendering. The consumeLodTreeUpdates path is intentionally unchanged — SparkRenderer already calls setDirty() via the lodDirty flow when it processes those updates inside a render frame.

Changes

• SplatPagerOptions: add optional onDirty?: () => void
• SplatPager: store callback, call it when processFetched() drains fetched to empty
• SparkRenderer: pass () => this.setDirty() when constructing the pager

Verified

Confirmed against an on-demand renderer that previously required a polling workaround to drive rendering during paged LOD loading. With this fix the workaround is no longer needed.

— Claude