Route fallback blocks through per-edge trampolines so the phi matches the CFG - #18

… the CFG

The LLVM backend emits invalid IR on Skyward Sword (SOUE01, Broadway). All 9917
chunk objects emit, then the module verifier rejects it:

  PHI node entries do not match predecessors!
    %fallback_pc = phi i32 [ %entry_pc, %cold_entry ],
                           [ -2143668204, %fallback_edge563 ], ...
    label %fallback_edge563
    label %indirect_taken562
  Instruction does not dominate all uses!   (x3, cascading, on %state.0)

emitter.cpp aliased every DOLIR_TERM_FALLBACK block onto the single shared
fallback block, whose first instruction is the fallback_pc phi. Only
fallbackDestination() and fallbackEdge() register an incoming value, so any code
branching to blocks_[i] -- notably the switch built for DOLIR_TERM_INDIRECT --
created a predecessor with no phi entry.

Hand out a per-edge trampoline instead of the raw shared block, so every edge
carries its own guest pc by construction rather than depending on each call site
remembering to register one. fallbackDestination() becomes a thin wrapper over
fallbackEdge(), which removes the fragile GetInsertBlock()-at-call-time logic.

Why only this title: GM4E01, GLME01 and GC6E01 all build clean on the same
backend. SOUE01 is 1,266,744 instructions -- 1.7x Mario Kart -- and Wii/Broadway
code carries far more indirect branches, so it is the first to exercise the path.

tests/llvm_fixture.cpp gains a case covering indirect dispatch into a fallback
continuation with modified state; it fails on the old emitter and passes here.

Verified: ctest 27/28 before and after on the same tree (the one failure was the
pre-existing module_abi AArch64 enum truncation, since fixed upstream by
528a204); Skyward Sword now emits 9917 objects with no verifier errors; Mario
Kart still emits 5803 clean.