Use Scratch.jl for GAP root

[fingolfin]

This also forces a full regeneration of our custom GAP root dir each time the package loads, instead of only during precompilation. This adds ~130 milliseconds startup overhead for me, of which 90% are spent in finding working C/C++ compilers... Not great, but with Scratch.jl any of the directories could move, so we kinda have to do it this way, as far as I can tell (Polymake.jl also does it this way). Perhaps I could at least cache the C/C++ compiler values somehow, but for now I just want to have this working.

Resolves #822
Resolves #728

[fingolfin]

This introduces a minor regression which I have not yet understood: upon using GAP, if GAP was already precompiled, all is fine; but if GAP gets precompiled then we don't show the GAP banner anymore. Huh?

[codecov]

Codecov Report

Merging #825 (cfd6e24) into master (df089c9) will increase coverage by 3.49%.
The diff coverage is 76.92%.

Additional details and impacted files

@@            Coverage Diff             @@
##           master     #825      +/-   ##
==========================================
+ Coverage   78.05%   81.54%   +3.49%     
==========================================
  Files          48       48              
  Lines        3686     3680       -6     
==========================================
+ Hits         2877     3001     +124     
+ Misses        809      679     -130     

Impacted Files	Coverage Δ
src/GAP.jl	82.35% <75.00%> (+0.17%)	⬆️
src/setup.jl	81.33% <75.00%> (+81.33%)	⬆️
src/packages.jl	64.15% <100.00%> (+0.68%)	⬆️

[ThomasBreuer]

@fingolfin I cannot reproduce the problem with the GAP banner.
I have checked out this branch and called Pkg.resolve().
After starting Julia and calling using GAP, the GAP banner is shown in my Julia session, no matter whether the message Info: Precompiling GAP is shown or not.

[ThomasBreuer]

Looks good. Thanks.

[fingolfin]

@ThomasBreuer which Julia version are you using? For me the problem of the missing banner only occurs in Julia 1.8 -- and indeed, the hack we use (I think perhaps invented by @thofma ?) checks for a function _tryrequire_from_serialized -- and starting with Julia 1.8 this function is called in more places. So I am guessing this code may need to be adapted to work for Julia 1.8.

[thofma]

Yes, this was my doing. I can have a look.

[thofma]

I would suggest you merge this here and I will fix it in a followup PR.