B018: report useless unary and arithmetic operations

[Eljees]

Refs #452.

B018 reported a statement that is only a literal, a tuple, or a call to a side-effect-free builtin. The case in the issue is an operation:

result = a * b
+c * d  # meant as the continuation of the line above

It was never reported. This adds the two node types @JelleZijlstra suggested there, UnaryOp and BinOp, with the limits his comment pointed to:

• only arithmetic operators count (+ - * / // % **). Shifts, bitwise operators and @ are left alone, because libraries overload them for their side effects: Airflow's task1 >> task2, and pipeline or query DSLs built on |. An expression that contains any of them anywhere is skipped;
• an operation whose operands call, await, yield or assign is skipped, since f() + 1 still runs f;
• an operation directly in a try or with block is skipped. That is where tests run one only to see it raise, as in with pytest.raises(TypeError): "1" + 1. tests/eval_files/b908.py contains exactly that, and it would otherwise start reporting.

An earlier attempt, #540, was closed the same day with no review. It also added Attribute, which this PR leaves out: a bare obj.attr can run a property getter for its side effect.

The message names the node like the other B018 cases: Found useless BinOp expression / UnaryOp.

Tests: tests/eval_files/b018_operators.py holds the 12 reported shapes (the issue's +c * d, each arithmetic operator, and -1, -a, +a, ~a, not a) and the cases left alone: overloaded operators, side effects in operands, and try/with blocks. On main it reports none of the 12. With the change, pytest tests passes on 3.12 (82 passed, 2 skipped) and on 3.10 (73 passed, 11 skipped). black --preview and flake8 are clean, and mypy bugbear.py reports the same 4 errors as on main. There is also a README note and a changelog entry under UNRELEASED.

[cooperlees]

This is muse (muse-spark), posting on behalf of @cooperlees via his coding agent.

Reviewed the diff and verified locally: checked out the PR head (29718a2, incl. the main-merge resolving the UNRELEASED changelog conflict with #578) and ran the full suite in an isolated venv in /tmp — 84 passed, and tests flagged no regressions in the neighboring B018/B908 eval files.

What I checked:

• The issue's case from #452 (result = a * b followed by orphaned +c * d) is now reported as a useless BinOp — good, this was a real gap in B018.
• Operator scoping is well-judged: only arithmetic (+ - * / // % **) count; shifts/bitwise/@ are excluded since libraries overload them for side effects (Airflow task1 >> task2, DSLs on |), and any non-arithmetic op anywhere in the expression tree vetoes the report.
• Side-effect guard in _is_useless_operation (Call/Await/Yield/YieldFrom/NamedExpr anywhere in operands) correctly skips f() + 1, -f(), (x := a) + 1.
• The try/with carve-out works: I ran flake8 directly against tests/eval_files/b908.py and confirmed no new B018 fires on the with raises: "1" + 1 lines, only the pre-existing B017s.
• Leaving out Attribute (unlike the closed #540) is right — a bare obj.attr can run a property getter.
• New eval file tests/eval_files/b018_operators.py fires exactly the 12 expected lines (verified by running flake8 directly: lines 10-21); README note and UNRELEASED changelog entry are clear and accurate.

No changes requested. Approving for merge.

[cooperlees]

I agree with AI. Thanks for this!