Flow-sensitive typing ignores dictionary key mutation

Bug Report

mypy allows runtime type checks for dictionary items. However it ignores subsequent dictionary mutations. This leads to mypy accepting code that fails at runtime.

To Reproduce

from typing import Dict, Optional

foo_dict: Dict[str, Optional[str]] = {'bar': 'baz'}

if foo_dict['bar'] is not None:
    foo_dict['bar'] = None
    foo_dict['bar'].upper()

Expected Behavior

mypy reports .upper() call as error. This is similar to what already happens with object attributes:

from dataclasses import dataclass
from typing import Optional

@dataclass
class Foo:
    bar: Optional[str]

foo = Foo(bar='baz')

if foo.bar is not None:
    foo.bar = None
    foo.bar.upper()

For such code mypy outputs the following error: "None" has no attribute "upper".

Actual Behavior

mypy reports no error.

Your Environment

• Mypy version used: 0.812
• Mypy command-line flags: --strict
• Mypy configuration options from mypy.ini (and other config files): none
• Python version used: 3.6.8
• Operating system and version: Windows 8.1

Start by running the Python reproduction with mypy --strict and compare it with the dataclass attribute example, confirming the dictionary case is accepted while the attribute case is rejected. Trace the flow-sensitive narrowing logic for dictionary item expressions and add coverage for the mutation; done when mypy reports the .upper() call as an error.