Optimize scan phase: in-memory change detection and batched DB writes

Performance improvements:
- Load all known files from DB into memory at startup
- Check file changes against in-memory map (no per-file DB queries)
- Batch database writes in groups of 1000 files per transaction
- Scan phase now only counts regular files, not directories

This should improve scan speed from ~600 files/sec to potentially
10,000+ files/sec by eliminating per-file database round trips.

internal/snapshot/file_change_test.go

@@ -194,8 +194,8 @@ func TestMultipleFileChanges(t *testing.T) {
 	// First scan
 	result1, err := scanner.Scan(ctx, "/", snapshotID1)
 	require.NoError(t, err)
-	// 4 files because root directory is also counted
-	assert.Equal(t, 4, result1.FilesScanned)
+	// Only regular files are counted, not directories
+	assert.Equal(t, 3, result1.FilesScanned)
 
 	// Modify two files
 	time.Sleep(10 * time.Millisecond) // Ensure mtime changes
@@ -221,9 +221,8 @@ func TestMultipleFileChanges(t *testing.T) {
 	result2, err := scanner.Scan(ctx, "/", snapshotID2)
 	require.NoError(t, err)
 
-	// The scanner might examine more items than just our files (includes directories, etc)
-	// We should verify that at least our expected files were scanned
-	assert.GreaterOrEqual(t, result2.FilesScanned, 4, "Should scan at least 4 files (3 files + root dir)")
+	// Only regular files are counted, not directories
+	assert.Equal(t, 3, result2.FilesScanned)
 
 	// Verify each file has exactly one set of chunks
 	for path := range files {