Fewer ByteString allocations in CPS parser with ByteStringIdx

[mzabani]

Part 2 of performance optimizations

Compared to master:

• Tuple Stream benchmark (with lots of runs, times of each run added together) went from 38433.4 to 36612.7ms, total memory allocated from 9470MB to 8410MB. That's a 4.8% wall time reduction and a 11.2% total memory allocated reduction
• Record stream benchmark went from 145.009s to 133.26s, total memory allocated from 10314MB to 9668MB. That's a 8.1% wall time reduction and a ~6.3% total memory allocated reduction

From the GHC Core, the main difference here is that in row decoders, parser continuations don't require allocating new ByteStrings, just Ints (ByteStringIdx). I don't expect this made a difference here, but it will be useful for future optimizations.

What this also does is push the part of the code that converts ByteString -> [a] (where a is the user's target row type) down in a way that we can avoid creating one instance of ByteString per row, and instead pass a big chunk of rows as the same single ByteString, and further down we can parseMany directly from that to [a].