refactor(assets): merge AssetInfo and AssetCacheState into AssetReference

This change solves the basename collision bug by using UNIQUE(file_path) on the
unified asset_references table. Key changes:

Database:
- Migration 0005 merges asset_cache_states and asset_infos into asset_references
- AssetReference now contains: cache state fields (file_path, mtime_ns, needs_verify,
  is_missing, enrichment_level) plus info fields (name, owner_id, preview_id, etc.)
- AssetReferenceMeta replaces AssetInfoMeta
- AssetReferenceTag replaces AssetInfoTag
- UNIQUE constraint on file_path prevents duplicate entries for same file

Code:
- New unified query module: asset_reference.py (replaces asset_info.py, cache_state.py)
- Updated scanner, seeder, and services to use AssetReference
- Updated API routes to use reference_id instead of asset_info_id

Tests:
- All 175 unit tests updated and passing
- Integration tests require server environment (not run here)

Amp-Thread-ID: https://ampcode.com/threads/T-019c4fe8-9dcb-75ce-bea8-ea786343a581
Co-authored-by: Amp <amp@ampcode.com>

app/assets/api/upload.py

@@ -95,7 +95,7 @@ async def parse_multipart_upload(
             file_client_name = (field.filename or "").strip()
 
             if provided_hash and provided_hash_exists is True:
-                # If client supplied a hash that we know exists, drain but do not write to disk
+                # Hash exists - drain file but don't write to disk
                 try:
                     while True:
                         chunk = await field.read_chunk(8 * 1024 * 1024)