Write rclone uploads under a temporary name and move them into place (closes #266)

The rclone backend wrote each object straight to its key, so killing an
upload to a local or sftp remote left a truncated object there. The next
backup found the key, skipped the upload and recorded a snapshot that
could not be restored.

On every remote with a server-side move, an object is now written under
a name ending in `.partial` and moved onto its key; listings skip such
names. Rclone's own copy also requires the remote's PartialUploads flag;
this does not, because hdfs shows a file while it is written without
setting it. Remotes without a move are written in place, as the README
now says.

Model: opus-5-5

internal/storage/file_atomic_test.go

@@ -14,6 +14,9 @@ import (
 // errStreamInterrupted stands in for an upload cut off mid-stream.
 var errStreamInterrupted = errors.New("connection reset mid-upload")
 
+// testBlobKey is a key laid out as a blob's key is.
+const testBlobKey = "blobs/aa/bb/aabbccddeeff"
+
 // failingReader yields its data once, then fails.
 type failingReader struct {
 	data []byte
@@ -43,7 +46,7 @@ func TestFileStorer_InterruptedWriteLeavesNoTrustedObject(t *testing.T) {
 	}
 
 	ctx := context.Background()
-	key := "blobs/aa/bb/aabbccddeeff"
+	key := testBlobKey
 
 	err = f.PutWithProgress(ctx, key, &failingReader{data: []byte("partial")}, 4096, nil)
 	if err == nil {
@@ -79,7 +82,7 @@ func TestFileStorer_ListSkipsPartialFiles(t *testing.T) {
 	}
 
 	ctx := context.Background()
-	realKey := "blobs/aa/bb/aabbccddeeff"
+	realKey := testBlobKey
 
 	err = f.Put(ctx, realKey, strings.NewReader("blob-bytes"))
 	if err != nil {