replace position vectors with bitset iterators (-23.58%)

[friendlymatthew]

This PR eliminates the intermediate Vec<u32> position arrays by iterating comma and newline bitsets directly during field extraction

extract_positions took 20% of decode time. Previously, decode extracted every set bit from the comma and newline bitsets into Vec<u32> position arrays, then extract_from_cache() indexed into these arrays

Profiling showed Vec::push inside extract_positions as a hot spot due to the per-element capacity checks + writes across positions.

Now, this PR adds a bitset iterator stored directly in the decoder. When fields are extracted, bitsets are walked on the fly using trailing_zeros and bit clear, computing positions from the u64 itself