Async CTX virtual texture tile loading

[juliendms]

Summary

Move CTX (virtual texture) tile loading off the render thread. Today, every cache miss in VirtualTexture::getTile synchronously reads a tile file, decodes it, and uploads it to GL inside the patch render loop -- so a pan that reveals a dozen new patches stalls the frame for a dozen load+upload roundtrips.

After this change:

• A single shared worker thread (TileLoader, in an anonymous namespace inside virtualtex.cpp) handles file I/O and image decode for all VirtualTexture instances in the process.
• The render thread performs bounded GL uploads (kMaxUploadsPerFrame = 4 per VirtualTexture per frame) from a per-owner result queue at the start of beginUsage().
• Tile gains a State enum (NotLoaded / Queued / Ready / Failed) that replaces the implicit tex == nullptr / loadFailed pair. State is read and written exclusively on the render thread; the worker treats Tile* as an opaque cookie that round-trips through the request/result pair unchanged.
• Cancellation is by id: each VirtualTexture registers a uint64_t with the loader on construction and deregisters on destruction. Worker results are delivered under the registry mutex, so a destroyed owner can never see a stale callback.

Design notes

• No public interface change. LODSphereMesh is untouched. getTile, beginUsage, endUsage, bind signatures are unchanged.
• No GL on the worker thread. Image::load is the only worker-side call; ImageTexture construction is exclusively in beginUsage.
• getTile becomes non-blocking. On a miss it enqueues a request (idempotent via the state check) and returns the deepest already-Ready ancestor. The first frame after a tile becomes visible may show a coarser ancestor; once the worker delivers and beginUsage uploads, subsequent frames show the finer tile.
• FIFO request queue, unbounded. No priority, no drop policy in V1.
• Owner registry under a mutex. The worker pushes the result while holding the same mutex unregisterOwner takes, so result-queue lifetime and registry membership are kept consistent.

User-visible effect

Smooth panning across high-LOD CTX surfaces. Freshly-revealed patches show a brief lower-LOD blur for a handful of frames instead of a per-pan freeze.

Disclaimer: The code in this PR was generated by AI (Anthropic Claude).