Precompile Redline trampolines and drop the bridge from the runners

[andreaTP]

Both runners created the Cranelift bridge in every machine to compile their ABI trampolines. The build-time compiler now emits them with the function bodies, and the runners only load and link:

• function bodies call memory.grow, memmove/memset and the table stub directly with the platform ABI, so their trampolines are gone
• entry trampolines take the function pointer as a parameter and are compiled once per function type
• import trampolines leave an 8-byte slot for the host stub address, which the runner fills before making the code executable; a shared table can call them with another instance's ctx, so they cannot read it from the ctx

NativeCode, written as .native format version 2, replaces byte[][], withCompilerFunction is removed, and an enforcer rule keeps the bridge and the compiler out of both runners' dependencies.

Fixes #202

[andreaTP]

@headius this should fix some of your observations, especially regarding upfront resources consumption:

on quickjs:

	main	this branch
time	358 ms	19 ms
Java heap allocated	157 MB	6.2 MB
first instance	1.1 s	0.48 s

[headius]

@andreaTP I can give a HEAD build a try later today with my JRuby + Prism integration branch. I assume this branch includes #210?

[andreaTP]

Yes 210 is included, but let me close also 203 and do a final review before trying things out, we'll get there!