GeezerVison

The refcount and uninit transforms pass all values, including temporaries, as the universe of the must-defined analysis (analyze_must_defined_regs). DefinedVisitor only generates and kills registers, and both callers only check registers (is_maybe_undefined in refcount, split_blocks_at_uninits in uninit), so the temporaries are never needed. However, a must analysis starts with a copy of the universe for every basic block, so time and memory were O(blocks * values).

This makes the refcount transform very slow for large functions, such as generated code or long module top levels. Benchmark: a function with n statements of the form

    a{i} = xs[{i % 7}] + {i}
    if a{i}:
        t = a{i}

timing insert_ref_count_opcodes alone (max RSS of the process):

Fix: leave everything except registers out of the universe. The analysis results for registers are the same, so the output doesn't change:

• The generated IR for the benchmark functions is identical.
• The generated C for mypy/checker.py, mypy/checkexpr.py, mypy/semanal.py and mypy/messages.py (29 MB) is identical to master.

For typical functions the gain is small (must-defined analysis on those four modules: 1.28s -> 1.14s), since they are short.

The remaining cost for large functions is still superlinear, because the must-defined sets contain O(registers) items per block and run_analysis builds a new set for every op. Using bitsets could fix that, but it's a bigger change, so I left it out of this PR.