test_macos on Github Actions is broken

[straight-shoota]

test_macos job clearly doesn't work as it is supposed to. It completes successfully even on a broken commit where every other job fails: https://github.com/straight-shoota/crystal/runs/1583371708

I'm not sure what exactly is happening with nix, but it surely doesn't report failure to the job runner.

[straight-shoota]

So I looked a bit into the setup of this job, and there seem to be a couple of problems:

1. In bin/ci build the command nix-shell --pure --run 'TZ=America/New_York make std_spec clean threads=1 junit_output=.junit/std_spec.xml' fails. But the step is not reported as a failure. I don't know why. Does nix-shell silently swallow the error? The CI script certainly errors for on_nix_shell false (Detect failures in Nix build environment #9776 fixed this).
2. The actual error is /nix/store/r0n4qzr800498l23cq80mg0k03j3aqms-cctools-binutils-darwin-wrapper-927.0.2/nix-support/setup-hook: line 138: ${role_pre}${cmd^^}=${cmd}: bad substitution. Looks like macos ships a really old bash version (3.2) and the substitution requires at least 4.0. So that should be easy to fix, just install a recent bash.
3. A minor issue: (a) The entire nix environment is only set up in the build stage. So a lot of dependencies are just installed there. I think this should already happen in prepare_build. (b) And ideally it would also be cached. But may that's already happening (there is some cache configuration). I'm not sure what's going on. It feels slow for being cached.

[straight-shoota]

I couldn't actually reproduce the original error on circle-ci's mac runner:

static:project distiller$ bash --version
GNU bash, version 3.2.57(1)-release (x86_64-apple-darwin18)
Copyright (C) 2007 Free Software Foundation, Inc.
static:project distiller$ CI_NIX_SHELL=true bin/ci build
nix-shell --pure --run 'TZ=America/New_York make std_spec clean threads=1 junit_output=.junit/std_spec.xml'
Using /nix/store/n55dgnwhnc8a6m2q9qhcymvyc0lx3wkg-llvm-10.0.0/bin/llvm-config [version=10.0.0]
...

But I suppose that doesn't really help anything when it still fails on GHA.

[straight-shoota]

#10480 fixes the original error (2) and moves nix initialization to prepare_build stage (3a).

Still need to figure out why the nix error didn't fail the job (1) and clarify if caching works correctly (3b).