Don't truncate file before detecting language

[jtibshirani]

Currently, we truncate a file's contents to 2048 bytes before passing it to
go-enry. I ran into a few cases where this is causing us to misclassify
files.

This PR removes the truncation. It should still be fine in terms of
performance, since go-enry is quite fast in general: ~1ms in my local
testing, even for large files. And we only run language detection if we plan to
index the file, which means we skip binary files and large files.

[jtibshirani]

I tried indexing sgtest/megarepo, and did not detect a difference in latency:

go run ./cmd/zoekt-index -parallelism 1 ../megarepo

Before: 7 min, 4 sec
After: 7 min, 14 sec

[keegancsmith]

nice idea comparing latency on megarepo.