Fixes an issue where the summaries for delete and assoc had multiple implementations that applied to each others' call sites.

We had code like this:

// Array.qll
private class DeleteUnknownSummary extends DeleteSummary {
    DeleteUnknownSummary() {
      this = "delete" and
      not exists(DataFlow::Content::getKnownElementIndex(mc.getArgument(0)))
    }
    ...
}

// String.qll
private class DeleteSummary extends SimpleSummarizedCallable {
  DeleteSummary() { this = ["delete", "delete_prefix", "delete_suffix"] + ["", "!"] }
}

The existence of DeleteSummary in String.qll means all method calls named delete are associated with the summary whose underlying string value is delete. This means DeleteUnknownSummary, if it exists, ends up applying to all calls named delete, regardless of the extra condition in its charpred.

This means that calls of form x.delete(:foo) would return all elements of x, not just its :foo element (because the call was treated as having an unknown key).

A similar situation occurred with AssocUnknownSummary in Hash.qll and AssocSummary in Array.qll.

This is a big enough footgun that I decided to add a new QL4QL query for it, which is part of this PR as well. In fact, that query is how I found the issue with assoc (the issue with delete was discovered by investigating a FP).

Evaluation looks reasonable. We lose some call edges and a few alerts - it's hard to verify at a glance whether these were correct, but I have a branch where some rather blatant FPs were causing by this issue and are fixed by this change.