Use FastPFOR in RLE/BP encoder

[gaborcsardi]

The new code is not templated, as we only ever encode uint32_t values currently.
It is much cleaner now, and the encoder itself is probably faster, but this does not really make write_parquet() faster, at least not for flights.

In fact, it is slightly slower for flights x 200, probably because we pack a lot of sections that are shorter than 32 values, so we need to copy them into a buffer first. I guess I can use the old bit packer for these shorter sections.

Bit-packing directly into the output buffer is also problematic because the bit packer is writing out unaligned uint32_t values. It seems to work on aarch64 and x86_64, but it might now work correctly everywhere.