Make parser LALR, conflict-free

[rhendric]

Motivation

This improves parsing performance by ~6%, maybe more.

The only reason GLR parsing was needed was to resolve a reduce-reduce conflict that I've factored away (along with the one shift-reduce conflict).

Context

Below is a plot of the time required to parse every Nix file in Nixpkgs, at each commit in this PR (100 runs each, median times are marked):

Note that benchmarks are based on nix-instantiate --parse, which includes the cost of showing expressions and sending them to /dev/null, so improvement may be more significant than presented.

I recommend reading this PR commit-by-commit. Each commit is hopefully self-explanatory, but the first three are only there so that the final GLR -> LALR refactoring goes smoothly. As you can see, they have no impact on parsing time.

In the first commit, I took the liberty of using Bison's named references feature for the rules I refactored, because it makes the refactoring less accident-prone and easier to read. I can back that out if it isn't house style, but I highly recommend it for anything but the simplest Bison grammars.

Priorities and Process

Add 👍 to pull requests you find important.

The Nix maintainer team uses a GitHub project board to schedule and track reviews.

[tomberek]

Tested and reviewed with @djacu and @JeremiahSecrist

Ran through tests and saw a perf improvement. Also ran into a segfault as stated here: #11141 (comment)

master is in a broken state at the moment, so let's defer this for the moment

[tomberek]

Deferring this change until after the next release. Otherwise, all changes look good. Only remaining question is to understand the intended behavior of:

#define YYLTYPE_IS_TRIVIAL 1

Would also like a rebase to re-run against the new CI regression checks.

[Mic92]

Good job!