Fix deferred compile trampoline using blr instead of br

Summary:
The AArch64 JitHelpers indirection stub for deferred function compilation used `blr` (branch with link) to call the failed_deferred_compile_trampoline, which clobbered the link register (x30) with the address of the next instruction (the resume entry point). When the trampoline returned, it jumped to the resume entry instead of returning to the original call site in the JIT-compiled function.

The x86 version correctly uses `jmp` (no return address push). Changed the AArch64 version to use `br` (branch without link) to match, preserving the original return address in x30 so the trampoline returns to the correct call site.

This fixes segfaults in async/await tests and other tests involving unjitable functions called from JIT-compiled code (e.g., test_awaited_invoke_function_unjitable).

Reviewed By: kddnewton

Differential Revision: D93875145

fbshipit-source-id: 98126def5763ad2a2103bffb118f76a0634d3546

Author: DinoV

cinderx/Jit/codegen/gen_asm.cpp

@@ -2176,7 +2176,7 @@ void NativeGenerator::generateEpilogue(BaseNode* epilogue_cursor) {
       as_->bind(trampoline);
       as_->mov(a64::x10, reinterpret_cast<uint64_t>(x.first));
       as_->mov(arch::reg_scratch_br, failed_deferred_compile_trampoline_);
-      as_->blr(arch::reg_scratch_br);
+      as_->br(arch::reg_scratch_br);
       x.second.trampoline = trampoline;
     }
     env_.addAnnotation("JitHelpers", jit_helpers);