[CodeCompletion] Avoid spurious signature help for init-list args

Somewhat surprisingly, signature help is emitted as a side-effect of
computing the expected type of a function argument.
The reason is that both actions require enumerating the possible
function signatures and running partial overload resolution, and doing
this twice would be wasteful and complicated.

Change #1: document this, it's subtle :-)

However, sometimes we need to compute the expected type without having
reached the code completion cursor yet - in particular to allow
completion of designators.
eb4ab335 did this but introduced a
regression - it emits signature help in the wrong location as a side-effect.

Change #2: only emit signature help if the code completion cursor was reached.

Currently there is PP.isCodeCompletionReached(), but we can't use it
because it's set *after* running code completion.
It'd be nice to set this implicitly when the completion token is lexed,
but ConsumeCodeCompletionToken() makes this complicated.

Change #3: call cutOffParsing() *first* when seeing a completion token.

After this, the fact that the Sema::Produce*SignatureHelp() functions
are even more confusing, as they only sometimes do that.
I don't want to rename them in this patch as it's another large
mechanical change, but we should soon.

Change #4: prepare to rename ProduceSignatureHelp() to GuessArgumentType() etc.

Differential Revision: https://reviews.llvm.org/D98488