[red-knot] Binary operator inference for union types

[sharkdp]

Summary

This fixes a bug I noticed while looking at ecosystem results. The MRE version of it is this:

def sub(x: float, y: float):
    # Red Knot: Operator `-` is unsupported between objects of type `int | float` and `int | float`
    return x - y

It's yet another instance of the general problem that our way of representing union types does not prevent us from handling unions (and intersections) incorrectly. The pattern is usually this: we implement operation X by explicitly matching on some cases that we handle explicitly (e.g. literals). For all other types, we fall back to some generic behavior (calling some functions that work for every Type). That generic behavior also compiles for Type::Union/Type::Intersection, but what we really want is to recursively call X for each element, which also gives us that special behavior for elements of the union.

For intersections, I'll open a ticket.

Test Plan

• New Markdown tests.
• Expected diff in the ecosystem checks

[github-actions]

mypy_primer results

Changes were detected when running on open source projects

mypy_primer (https://github.com/hauntsaninja/mypy_primer)
- 
-     |
-     |
- error: lint:unsupported-operator
-    --> /tmp/mypy_primer/projects/mypy_primer/mypy_primer/utils.py:114:79
- 112 |     if check and proc.returncode:
- 113 |         raise subprocess.CalledProcessError(proc.returncode, cmd, output=stdout, stderr=stderr)
- 114 |     return subprocess.CompletedProcess(cmd, proc.returncode, stdout, stderr), end_t - start_t
-     |                                                                               ^^^^^^^^^^^^^^^ Operator `-` is unsupported between objects of type `int | float` and `int | float`
- Found 50 diagnostics
+ Found 49 diagnostics

[sharkdp]

Instead of short-circuiting here, we could probably also build a better diagnostic, telling you which exact element-combination didn't work. But it seems low priority and doesn't necessarily have to be included here, I think.

[carljm]

Thank you!