MemoryIndex and FVH don't play along with multi-value fields [LUCENE-5011]

When multi-value fields are indexed to a MemoryIndex, positions are computed correctly on search but the start and end offsets and the values array index aren't correct.

Comparing the same execution path for IndexReader on a Directory impl and MemoryIndex (same document, same query, same analyzer, different Index impl), the difference first shows in FieldTermStack.java line 125:

termList.add( new TermInfo( term, dpEnum.startOffset(), dpEnum.endOffset(), pos, weight ) );

dpEnum.startOffset() and dpEnum.endOffset don't match between implementations.

This looks like a bug in MemoryIndex, which doesn't seem to handle tokenized multi-value fields all too well when positions and offsets are required.

I should also mention we are using an Analyzer which outputs several tokens at a position (a la SynonymFilter), but I don't believe this is related.

---
Migrated from [LUCENE-5011](https://issues.apache.org/jira/browse/LUCENE-5011) by Itamar Syn-Hershko, updated Jul 31 2014

Start at FieldTermStack.java line 125 and compare the same document, query, analyzer, and token stream through Directory and MemoryIndex. Inspect MemoryIndex handling of tokenized multi-value fields, positions, offsets, and values-array indexes; done means the implementations produce matching offsets and value indexes while preserving the currently correct positions.

Written by the indexing model from the issue text.