Fix to avoid locale-dependent string comparisons

It doesn’t seem appropriate to use locale dependent ordering to sort the
variants and extensions for normalization. 
The result of normalization should be consistent.

This commit removes a dependency on the ECMAScript internationalization API.
This API is not available everywhere, particularly on embedded devices like
TVs where storage space for internationalization data is
limited.

The new test case isn’t strictly valid - there should not be more than one
extension with the same singleton.

Closes GH-3.

lib/index.js

@@ -29,8 +29,6 @@ import {likely} from './likely.js'
 
 const own = {}.hasOwnProperty
 
-const collator = new Intl.Collator()
-
 /**
  * @param {Schema} base
  * @param {Partial<Schema>} changes
@@ -159,7 +157,7 @@ export function bcp47Normalize(tag, options) {
   removeLikelySubtags(schema)
 
   // 5. Sort variants, and sort extensions on singleton.
-  schema.variants.sort(collator.compare)
+  schema.variants.sort()
   schema.extensions.sort(compareSingleton)
 
   // 6. Warn if fields (currently only regions) should be updated but have
@@ -323,5 +321,15 @@ function add(object, key, value) {
  * @returns {number}
  */
 function compareSingleton(left, right) {
-  return collator.compare(left.singleton, right.singleton)
+  if (left.singleton > right.singleton) {
+    return 1
+  }
+
+  if (left.singleton < right.singleton) {
+    return -1
+  }
+
+  // It is invalid to have more than one extension with the same singleton so
+  // we should never reach this code.
+  return 0
 }

test.js

@@ -220,6 +220,7 @@ test('fixtures', function () {
     'en-us': 'en',
     'en-gb': 'en-GB',
     'en-us-x-twain': 'en-x-twain',
+    'en-a-ext1-a-ext2': 'en-a-ext1-a-ext2',
     'en-a-myext-b-another': 'en-a-myext-b-another',
     'en-bu': 'en-MM',
     en: 'en',