strictly better -ODebug machine code quality than LLVM #24144

This is a meta issue to track enhancements to x86_64 machine code quality. Any example Zig code for which -ODebug -fllvm produces better x86_64 code than -ODebug -fno-llvm can be reported as a sub-task of this one.

• codegen quality: unnecessary function prologue and epilogue #22585
• switch branch trees into partial jump tables (issue needs a writeup)
• continue with comptime-known operand (issue needs a writeup)
• x86_64 backend fails to reuse temporary stack values resulting in greater stack usage and potential overflow #24183
• restore memory offset tracking (issue needs a writeup)
• implement relocations for arbitrary x86_64 instructions in the linker (issue needs a writeup)
• self-hosted compiler generates code that is too slow (~500x slower than -fllvm -O Debug) #25111