Always load source files in binary mode.

Source files can be in any coding system, provided Python can
read a coding: line at the beginning of the file. So source files
should be loaded in binary format and decoded according to that
line, not assumed to be in the default coding system.

Fixes #398.

test/test_regression.py

@@ -4,6 +4,7 @@ found a good place in any other testing module.
 """
 
 import os
+import sys
 import textwrap
 
 from .helpers import TestCase, cwd_at
@@ -154,3 +155,19 @@ class TestRegression(TestCase):
             "    yield 1\n" \
             "abc()."
         assert Script(s).completions()
+
+
+def test_loading_unicode_files_with_bad_global_charset(monkeypatch, tmpdir):
+    dirname = str(tmpdir.mkdir('jedi-test'))
+    filename1 = os.path.join(dirname, 'test1.py')
+    filename2 = os.path.join(dirname, 'test2.py')
+    if sys.version_info < (3, 0):
+        data = "# coding: latin-1\nfoo = 'm\xf6p'\n"
+    else:
+        data = "# coding: latin-1\nfoo = 'm\xf6p'\n".encode("latin-1")
+
+    with open(filename1, "wb") as f:
+        f.write(data)
+    s = Script("from test1 import foo\nfoo.",
+               line=2, column=4, path=filename2)
+    s.complete()