Optimize gtf by only parsing lazily per-refName

[cmdcolin]

Possible fix for #2681

loading a random location on chr1 for a 19MB gtf.gz file from UCSC

This branch, ~10s
Main branch, ~48s

approx tracing differences

this branch

main branch

[cmdcolin]

Note that the 19mb gzip gtf unzips to 329mb of unzipped text so it ends up being kinda big data.

it actually comes close to the maximum string size in chrome (512MB). if we need to handle bigger than 512 then we would want to be careful about not converting the buffer directly to string via TextDecoder (silly blogpost I made on subject awhile back https://cmdcolin.github.io/posts/2021-10-30-spooky)