codegen: LLVM 21 ThreadSafeContext and EHFrame compatibility

[gbaraldi]

Summary

• Add withContextDo compatibility shim for ThreadSafeContext API changes in LLVM 21 (which removes getLock(), Lock, and getContext() in favor of callback-based withContextDo)
• Replace all manual lock/unlock patterns with scoped withContextDo callbacks across jitlayers.h, jitlayers.cpp, aotcompile.cpp, and disasm.cpp
• Add JLEHFrameRegistrationPlugin for LLVM 21+ (which removes jitlink::EHFrameRegistrar), using JITLink allocation actions for EH frame registration
• Debug asserts ensure get_context()/get_module() are only called inside withContextDo scope
• Lock hierarchy respected: TSC locks released before addOutput()/R->replace() which can trigger materialization

Supersedes #59946 and #60358, forward-porting those changes to current master.

Test plan

• Builds and passes smoke test on LLVM 20
• Builds and passes smoke test on LLVM 21 (tested with libLLVM 21.1.2 JLL)
• CI

This pull request was written with the assistance of generative AI (Claude).

🤖 Generated with Claude Code

There is quite a few whitespace changes so ignoring it is quite helpful

[giordano]

Analyzegc is unhappy: https://buildkite.com/julialang/julia-master/builds/55503#019cd3d0-ac7c-40b7-99ec-70394e193ab7/L2196

/cache/build/builder-amdci5-5/julialang/julia-master/src/jitlayers.cpp:984:9: error: Safepoints disabled at end of function [julia.GCChecker]
  984 |         F->setName(*Sym);
      |         ^
[...]
/cache/build/builder-amdci5-5/julialang/julia-master/src/jitlayers.cpp:2413:12: error: Calling potential safepoint as SimpleFunctionCall from function annotated JL_NOTSAFEPOINT [julia.GCChecker]
 2413 |     return withContextDo(TSM.getContext(), [&](LLVMContext*) -> std::unique_ptr<MemoryBuffer> {
      |            ^

[xal-0]

It's awkward that jl_codegen_output_t owns a ThreadSafeModule to begin with, especially since we have usually unique ownership of the llvm::Module and LLVMContext. I've got a branch that refactors it to hold a reference to the module, so we can just create the module and context directly in the typical case and .withModuleDo() to emit to a user-provided ThreadSafeModule. Could you drop c0284d5 and take 27715b2 from my branch instead?

[xal-0]

I deleted JLEHFrameRegistrationPlugin; we can use the upstream EHFrameRegistrationPlugin and pass our alloc actions directly.

[giordano]

All green. Is this ready?

[gbaraldi]

I think so thanks @xal-0 for pushing it forward

[oscardssmith]

Accidentally squashed. Sorry. Can we make it so that the dont squash label actually prevents squashing? The big green button is way too easy to press.