DefaultSolrHighlighter can sometimes avoid CachingTokenFilter

The DefaultSolrHighlighter (the most accurate one) is a bit over-eager to wrap the token stream in a CachingTokenFilter when hl.usePhraseHighlighter=true. This wastes memory, and it interferes with other optimizations – LUCENE-6034. Furthermore, the internal TermOffsetsTokenStream (used when TermVectors are used with this) wasn't properly delegating reset().