When calculating the MAC for a message using olm_sas_calculate_mac() and
olm_sas_calculate_mac_long_kdf() the resulting MAC will be base64
encoded using _olm_encode_base64().

The _olm_encode_base64() method requires an input buffer and output
buffer to be passed alongside the input length. The method is called
with the same buffer, containing the MAC, for the input buffer as well
as for the output buffer. This results in an incorrectly base64 encoded
MAC.

For example the byte array:
    [121, 105, 187, 19, 37, 94, 119, 248, 224, 34, 94, 29, 157, 5,
     15, 230, 246, 115, 236, 217, 80, 78, 56, 200, 80, 200, 82, 158,
     168, 179, 10, 230]

will be encoded as  eWm7NyVeVmXgbVhnYlZobllsWm9ibGxzV205aWJHeHo
instead of as       eWm7EyVed/jgIl4dnQUP5vZz7NlQTjjIUMhSnqizCuY

Notice the different value at the 10th character.

The correct result can be independently checked using Python for
example:

>>> from base64 import b64encode
>>> mac = [121, 105, 187, 19, 37, 94, 119, 248, 224, 34, 94, 29, 157, \
           5, 15, 230, 246, 115, 236, 217, 80, 78, 56, 200, 80, 200, \
           82, 158, 168, 179, 10, 230]
>>> b64encode(bytearray(mac)).rstrip(b"=")
>>> b'eWm7EyVed/jgIl4dnQUP5vZz7NlQTjjIUMhSnqizCuY'

This happens because the encode_base64() method that is used does not support
in-place encoding in the general case. This is because the remainder for a 32
bit input will always be 2 (32 % 6 == 2).

The remainder will be used over here:
https://gitlab.matrix.org/matrix-org/olm/-/blob/c01164f001d57fbe2297fe11954b58077a68dc0d/src/base64.cpp#L74

The logic that gets executed if a remainder exists depends on the original input
values, since those already got in-place encoded the whole block will behave
differently if the input buffer is the same as the output buffer.