Add hidden --overwrite-union-syntax option

[cdce8p]

Start updating the test output (errors and reveal type) to use the PEP 604 union syntax. The process is similar to what we did last year when we adopted PEP 585 generics in the test output.

As we still default to Python 3.9 for type checking, this PR adds a temporary internal flag --overwrite-union-syntax to be able to overwrite the output.

This PR only changes a handful of test files so the modifications to mypy itself are clearly visible. The goal is to change the test output for all other tests in followup PRs. Once that's done, I plan to deprecate the --force-union-syntax flag similar to the already deprecated --force-uppercase-builtins.

[bzoracler]

This is peripherally relevant, but has there been previous discussions about the union syntax being rather ambiguous for def callables in reveal_type?

def f1() -> int | None: return 1
def f2() -> int: return 1

reveal_type(f1)                     # Revealed type is "def () -> builtins.int | None"
reveal_type(f2 if int() else None)  # Revealed type is "def () -> builtins.int | None"

[cdce8p]

This is peripherally relevant, but has there been previous discussions about the union syntax being rather ambiguous for def callables in reveal_type?

Though ultimately rejected, PEP 677 discusses that here: https://peps.python.org/pep-0677/#making-bind-tighter-than.

[cdce8p]

def f1() -> int | None: return 1
def f2() -> int: return 1

reveal_type(f1)                     # Revealed type is "def () -> builtins.int | None"
reveal_type(f2 if int() else None)  # Revealed type is "def () -> builtins.int | None"

We should probably consider adding parentheses to the second example. As a point of comparison, pyright infers it as "(() -> int) | None".

[github-actions]

According to mypy_primer, this change doesn't affect type check results on a corpus of open source code. ✅

[cdce8p]

We should probably consider adding parentheses to the second example. As a point of comparison, pyright infers it as "(() -> int) | None".

I'd suggest we fix this after migrating the test output. That way would could immediately see the impact of a potential change. The fix should be a fairly simple change in the TypeStrVisitor.