refactor: Reduce intermediate allocations in encoding logic.

There were a few sources of memory inefficiency in the encode function: * converting everything to an intermediate list rather than working with the vector. * creating a builder, converting it to a bytestring, just to make it a builder again. This PR removes those allocations and also takes inspiration on efficient bytestring building from [hsthrift](https://github.com/facebookincubator/hsthrift/blob/905533410f36befd4672bc388c9821a2fadd8d65/lib/Thrift/Protocol/Binary.hs#L90). Benchmarks: 20% improvement in encoding. had to fix the benchmark to actually be meaningful for encode since it was always returning immediately. Also augmented the presidents dataset by replicating the columns until we had 20k of them to make sure the improvements weren't noise. Before: ``` benchmarking positional/encode/presidents/with conversion time 15.21 ms (14.52 ms .. 15.82 ms) 0.989 R² (0.983 R² .. 0.995 R²) mean 13.57 ms (13.26 ms .. 14.00 ms) std dev 943.0 μs (723.5 μs .. 1.047 ms) variance introduced by outliers: 33% (moderately inflated) benchmarking named/encode/presidents/with conversion time 22.10 ms (21.98 ms .. 22.22 ms) 1.000 R² (0.999 R² .. 1.000 R²) mean 22.15 ms (22.04 ms .. 22.29 ms) std dev 280.2 μs (204.9 μs .. 370.6 μs) ``` After: ``` benchmarking positional/encode/presidents/with conversion time 13.13 ms (13.07 ms .. 13.23 ms) 1.000 R² (1.000 R² .. 1.000 R²) mean 13.14 ms (13.10 ms .. 13.18 ms) std dev 110.9 μs (83.62 μs .. 159.7 μs) benchmarking named/encode/presidents/with conversion time 19.80 ms (19.62 ms .. 19.95 ms) 1.000 R² (0.999 R² .. 1.000 R²) mean 20.15 ms (20.02 ms .. 20.47 ms) std dev 416.2 μs (173.1 μs .. 772.1 μs) ```