Header names should be compared binary but not linguistic.

Header names should be compared binary but not linguistic.
Moreover according to RFC2616 header names are subset of ASCII chars [0-127],
so culture-aware comparison is useless and eats too many processor time
(according to my benchmarks get an item from Hashtable with OrdinalIgnoreCase
comparers more than 20x times faster compared to Hashtable with InvariantCultureIgnoreCase
comparer)