Improve performance of instruction_mode_to_global_mode pass

[vosen]

Currently, "instruction_mode_to_global_mode" is extremely slow. Here's a timing for a mildly big module (not the worst one):

Pass "normalize_identifiers" took 59.860832ms
Pass "replace_known_functions" took 2.624µs
Pass "normalize_predicates2" took 23.391069ms
Pass "resolve_function_pointers" took 7.802311ms
Pass "fix_special_registers" took 22.046943ms
Pass "expand_operands" took 22.284991ms
Pass "insert_post_saturation" took 6.117475ms
Pass "deparamize_functions" took 4.783183ms
Pass "replace_instructions_with_functions_fp_required" took 11.614769ms
Pass "normalize_basic_blocks" took 8.002911ms
Pass "remove_unreachable_basic_blocks" took 6.165235ms
Pass "instruction_mode_to_global_mode" took 21.265026257s
Pass "insert_explicit_load_store" took 92.962535ms
Pass "insert_implicit_conversions2" took 99.951435ms
Pass "replace_instructions_with_functions" took 73.920287ms
Pass "hoist_globals" took 5.73405ms
Pass "emit_llvm" took 260.021649ms
Wrote to module_0001_01_06.ll
Wrote to module_0001_01_06.linked.ll
Wrote to module_0001_01_06.elf
Wrote to module_0001_01_06.asm
Pass "compile_bitcode" took 13.447234329s

This pass is the most complicated of the ones we currently have, but there are some low-hanging performance fruits