Use Custom Rust Targets for Shared Wasm Libraries

[luc-blaeser]

Simplify Motoko RTS building:

• Use custom Rust targets for generating PIC-relocatable Wasm.
• Get rid of emscripten or explicit compilation via LLVM IR.
• Get rid of musl (until now, it was only used for classical persistence).
• Optimization: Use bulk-memory operations because memcpy of Rust is slower than musl.
• No need for the wasm-ld fix as the custom targets do not generate corresponding DSO-local relative table index relocations ([lld][WebAssembly] Fix relocation of Wasm table index with GOT access llvm/llvm-project#104043).

[github-actions]

Comparing from a6d1305 to 0b196f8:
In terms of gas, 5 tests regressed and the mean change is +0.0%.
In terms of size, 5 tests improved and the mean change is -41.1%.

[crusso]

I won't pretend to know what is going one but this seems like a nice cleanup removing musl from the classical compiler. Shame the instruction counts seem to go up.

[crusso]

I wonder why this has increased so much?

[luc-blaeser]

Good point. I need to analyze why.

[luc-blaeser]

I analyzed it and the problem is that the Rust/LLVM builtin memcpy is way slower than musl, not using 64-bit copying. I activated bulk memory copying to get better performance.
(Bignum needs a lot of memcpy because of big integer reallocations to larger blobs.)