Remove an in-memory buffer for LSIF transformation

Right now, we buffer the entire output of the LSIF transformer in memory before sending it on to gitlab, which is inefficient.

This MR allows us to stream the output to rails instead.

We have a benchmark for this path. Here's the before:

BenchmarkGenerate-8   	       1	2252189327 ns/op	         7.14 MiB/op

And here's the after:

BenchmarkGenerate-8   	       1	2432998654 ns/op	         2.25 MiB/op

Closes #284 (closed)