Fixed issue in emubd where noop erases start out-of-sync with disk

Because reproducibility is extremely important, emubd always zeros
blocks on the first erase, even when erase_value=-1.

Well, at least it should be. We were correctly zeroing the blocks in
RAM, but if erase_value=-1 we were leaving the disk unzeroed, causing
the disk to fall out of sync.

Fixed by zeroing disk in lfs_emubd_createcfg, even if erase_value=-1.

Also I went ahead and dropped the bd->disk->scratch block. We're already
allocating RAM-backed blocks on erase anyways, so keeping scratch around
doesn't really gain us anything anymore. Now there is just a temporary
allocation in lfs_emubd_createcfg to zero the disk efficiently during
initialization.

bd/lfs_emubd.h

@@ -67,9 +67,8 @@ typedef int64_t lfs_emubd_ssleep_t;
 
 // emubd config, this is required for testing
 struct lfs_emubd_config {
-    // 8-bit erase value to use for simulating erases. -1 does not simulate
-    // erases, which can speed up testing by avoiding the extra block-device
-    // operations to store the erase value.
+    // 8-bit erase value to use for simulating erases. -1 simulates a noop
+    // erase, which is faster than simulating a fixed erase value.
     int32_t erase_value;
 
     // Number of erase cycles before a block becomes "bad". The exact behavior
@@ -127,7 +126,6 @@ typedef struct lfs_emubd_block {
 typedef struct lfs_emubd_disk {
     uint32_t rc;
     int fd;
-    uint8_t *scratch;
 } lfs_emubd_disk_t;
 
 // emubd state