Do not do rereading step in fread

[st-pasha]

Since the very early days, the design of fread was such that if any "type-bumps" happen in the middle of a file, then the corresponding column(s) will be marked as "require re-reading", and then at the end we would re-parse the entire file using the new column type. This approach has obvious drawbacks:

• the amount of time necessary to read the file almost doubles (and type bumps happen more often than you'd think);
• the logic behind type-bumping is very error-prone;
• it is impossible to read a stream-like input, where the data cannot be arbitrarily rewound;

The new approach for handling type-bumping is the following:

• When a type-bump occurs while reading a chunk, we enter the ordered section and temporarily suspend execution of other threads;
• While all other threads are paused, we:
  • "archive" the column which was type-bumped;
  • update the global types array with the new column parse types;
• After that, the parallel execution resumes from the start of the type-bumped chunk;
• (the process above may occur multiple times with different columns or different chunks);
• In the end, when the final frame is constructed, the columns that were comprised of multiple chunks are rbound together with automatic type promotion.

Closes #1843
Closes #1446

[oleksiyskononenko]

If we remove ParserIterable ParserLibrary::successor_types(dt::read::PT pt) const; method (that is not used anywhere anyways) ParserIterable and ParserIterator classes could be safely removed too.

[st-pasha]

Yes, we need to do more cleanup around the ParserLibrary in general, but it's better to defer to a future PR.