Fix FAT16 detection

SD_t3.h

 // This Teensy 3.x optimized version is a work-in-progress.
 // Uncomment this line to use the Teensy version.  Otherwise,
 // the normal SD library is used.
-#define USE_TEENSY3_OPTIMIZED_CODE
+//#define USE_TEENSY3_OPTIMIZED_CODE
 
 /* Why reinvent the SD library wheel...
  *   1: Allow reading files from within interrupts

cache_t3.cpp

 #define CACHE_FLAG_IS_DIRTY  2
 #define CACHE_FLAG_IS_FAT    4
 
-/*
+//#define PRINT_SECTORS
+
+#ifdef PRINT_SECTORS
 static void print_sector(const void *data)
 {
 	const uint8_t *p = (const uint8_t *)data;
 		if ((i & 31) == 31) Serial.println();
 	}
 }
-*/
+#endif
 
 // Read a sector into the cache.  If the sector is already cached,
 // of course no actual read occurs.  This is the primary function
 				c->flags = CACHE_FLAG_HAS_DATA;
 				if (is_fat) c->flags |= CACHE_FLAG_IS_FAT;
 				ret = &c->data;
-				 //Serial.printf("cache read %u\n", lba);
-				 //print_sector(&c->data);
+				 #ifdef PRINT_SECTORS
+				 Serial.printf("cache read %u\n", lba);
+				 print_sector(&c->data);
+				 #endif
 			}
 		}
 	}

init_t3.cpp

 	//Serial.printf(" sectors_per_fat = %d\n", sectors_per_fat);
 	uint32_t root_dir_entries = unaligned_read16_align8(vol->u8 + BPB_RootEntCnt);
 	//Serial.printf(" root_dir_entries = %d\n", root_dir_entries);
-	uint32_t root_dir_sectors = (root_dir_entries + 31) >> 5;
+	uint32_t root_dir_sectors = (root_dir_entries + 15) >> 4;
 	//Serial.printf(" root_dir_sectors = %d\n", root_dir_sectors);
 
 	uint32_t total_sectors = unaligned_read16_align8(vol->u8 + BPB_TotSec16);
 
 	fat1_begin_lba = partition_lba + reserved_sectors;
 	fat2_begin_lba = fat1_begin_lba + sectors_per_fat;
-	data_begin_lba = fat2_begin_lba + sectors_per_fat;
+	data_begin_lba = fat2_begin_lba + sectors_per_fat + root_dir_sectors;
 
 	uint32_t cluster_count = (total_sectors - reserved_sectors
 		- root_dir_sectors - (sectors_per_fat << 1)) >> s2c;
 	} else if (cluster_count < 65525) {
 		fat_type = 16;
 		rootDir.length = root_dir_entries << 5;
-		rootDir.start_cluster = partition_lba + reserved_sectors;
+		rootDir.start_cluster = partition_lba + reserved_sectors + sectors_per_fat * 2;
 		rootDir.type = FILE_DIR_ROOT16;
 	} else {
 		fat_type = 32;