Improved checking of nullable operands in expressions

[ahejlsberg]

This PR improves our checking of nullable operands in expressions. Specifically, we now error in the following cases:

• If either operand of a + operator is nullable, and neither operand is of type any or string.
• If either operand of a -, *, **, /, %, <<, >>, >>>, &, |, or ^ operator is nullable.
• If either operand of a <, >, <=, >=, or in operator is nullable.
• If the right operand of an instanceof operator is nullable.
• If the operand of a +, -, ~, ++, or -- unary operator is nullable.

An operand is considered nullable if the type of the operand is null or undefined or a union type that includes null or undefined. Note that the union type case only only occurs in --strictNullChecks mode because null and undefined disappear from unions in classic type checking mode.

Fixes #12795.

[DanielRosenwasser]

Object is definitely null, and typically we don't think of objects as being operands to arithmetic operations (whereas we do when it comes to primitives). Would be nice if we could be more precise on the error message.

[ahejlsberg]

Two orthogonal issues:

First, we could do extra work to omit the word possibly when the type is exactly null or undefined. However, it is very rare for someone to explicitly use null or undefined as an operand, so it's not clear that it really matters all that much. This is the type of error that only surfaces in a test.

Second, we could consider using the word value instead of object when reporting for expression operands--but it's really the same error that we use for obj.foo where obj may be null or undefined. The downside would be error message searchability, but since the error is pretty obvious that may not be an issue.