Only define the type name table in the main LLVM module

[HertzDevil]

#15904 made every LLVM module that references the #as pseudo-method define its own copy of __crystal_type_id_to_class_name_map, when only one copy in the main LLVM module is sufficient. This PR brings the total cache size of the compiler itself from 234 MB to 148 MB on my Windows machine, for example.

[miry]

TLDR: Thank you very much, @HertzDevil — it works on my machine! 🎉

Since v1.17.0, GHA could no longer compile my application — same for nightly builds.
Locally, I also noticed performance degradation, especially after clearing the cache or using --release.

The last stable version that worked fine is v1.16.3:

• Cache folder: 428 MB
• Build time without cache: 4:50 min
• Build time with cache: 1:35 min

I built Crystal 1.17.0 with LLVM 20 and saw much higher disk writes and memory usage.
Swap grew from 0 GB to 8 GB (I have 16 GB RAM).

• Cache folder: 1.4 GB
• Build time without cache: 6:35 min
• Build time with cache: 5:44 min

The amount of files in the cache folder (~8800) did not change,
but most files increased in size from around 200 KB to 2 MB.

This PR significantly improves the stats:

• Cach folder: 426 MB 🎉
• Build time without cache: 3:30 min
• Build time with cache: 1:30 min

Amazing work by @HertzDevil — super fast at finding and fixing the issue! 🙌

---

My learning section from @HertzDevil:

1. llvm-dis makes bs files human-readable:

   llvm-dis M-arten5858D-B-58-01956c356b7343c30982cb67961c1e27.o0.bc -o - > v1.17.0_dis.txt
   

2. Using diff helps spot codegen changes, e.g.:

    < @"'str'.10876" = private constant ... c"Tuple(Array(AllowlistedJWT), Array(AllowlistedJWT)).class\00"
    < @"'str'.10877" = private constant ... c"Tuple(AllowlistedJWT, Bool | Float32 | Float64 | ...).class\00"

for me, AllowlistedJWT occurrence count jumped from 25 → 115.