The old implementation allocated a new string for every invocation, and
repeatedly scanned the string for occurences of the various Windows device
names. This commits resizes the original string instead if needed, and
detects all devices with a single pass.
Every single call to base64decode allocates a mapping table for all
base64 characters. This is quite wasteful, as the map is in fact static.
We could use a static variable here, but that would have unpleasant consequences
if we ever encounter input with non-valid base64 characters (which
implicitly modifies the map).
The number of character ranges for base64 is quite low (3, plus 4 exceptions),
thus we can simply check that explicitly in code instead of using a dynamic hash table.
