Fix base64 decoding for length= argument and non-padding decoding

When decoding only a substring, we naturally see no padding, so we must
not base tail processing on the number of pad characters seen, but
rather on the number of characters missing until the end of the current
block of four.

Fixes #3378

bin/varnishtest/tests/m00042.vtc

@@ -263,7 +263,7 @@ client c1 {
 	expect resp.http.hexmix2hexlc == resp.http.hexmix2hex
 	expect resp.http.hexparam == "0123456789"
 	expect resp.http.b642b64 == "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefghij"
-	expect resp.http.b64url2b64url == "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefeQ=="
+	expect resp.http.b64url2b64url == "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefgg=="
 	expect resp.http.b64urlnopad2b64urlnopad == "ABCDEFGHIJKLMNOPQRSTUVWXYZabcdefgg"
 } -run
 

lib/libvmod_blob/base64.c

@@ -185,7 +185,7 @@ base64_decode(const enum encoding dec, char *restrict const buf,
 		}
 	}
 	if (n) {
-		if (!alpha->padding)
+		if (n - term != 0)
 			u <<= (6 * (4 - n));
 		if (decode(&dest, buf, buflen, u, n-term) < 0)
 			return -1;