Clarify rules on safety and sizeof

[BillWagner]

This is a small focused change, but fixes what I (at least) misinterpreted at least twice.

In version 7.3 we correctly updated the rules for in 8.8 for Unmanaged types to include struct types where the fields are all unmanaged types. This includes generic types where the type parameter has the unmanaged constraint.

However, we didn't restate the restrictions for sizeof in 12.8.19 to prevent extending the allowed types in safe context to include those types.

I updated the language in 12.8.19 to make those restrictions clear.

[Nigel-Ecma]

LGTM

§24.6.9 is as undefined as ever, but that is by design.

[jskeet]

Adding as a comment here so we can discuss it... 24.6.9 starts with:

For certain predefined types (§12.8.19), the sizeof operator yields a constant int value. For all other types, the result of the sizeof operator is implementation-defined and is classified as a value, not a constant.

It's not clear from that sentence that sizeof yields a constant int value for enums. Should we update that as well? (I'm happy to create a PR for that.)

Separately, 24.6.9 doesn't have a restriction on sizeof only being used for unmanaged types, so there's nothing to stop sizeof(MyClassType) being valid in an unsafe context - it's implementation-specific. Is that the intention?

[jskeet]

Bill will remove "predefined" from the first sentence in 24.6.9 and add the line break as suggested, then we can merge.