add a --flat mode, maybe make it the default

[nikomatsakis]

I was thinking that a good default mode would be --flat, which would be in contrast to the tree modes, and would work as follows.

The LEAF data structure is a mapping like

struct Leaves {
    /// key: name of top frame
    map: HashMap<String, Leaf>
}

struct Leaf {
    /// increment this once for every stack trace with this function as the top
    /// of the stack. This will be equal to `contexts.values().sum()`, so maybe
    /// we don't need it as an explicit value.
    total_hits: usize,

    /// each time we encounter this leaf, we take the remaining stack frames
    /// (i.e., its callers) and we put them in this map with an initial value of 1;
    /// if the same callers are already found, then we just bump the usize value.
    ///
    /// This way, we have an idea of where the leaf was called from and how many
    /// times from each trace.
    contexts: HashMap<Vec<String>, usize>,
}

• We populate the data structure above initially based on all the frames
• Then we go over all the leaves:
  • for each one that represents less than N% of total running time:
  • we pop the leaf and go to each of the contexts:
    • pop the next top-most frame from the context, and account the time for this leaf over into that context

We keep doing this until all leaves have N%

Then we print the leaves, in order of decreasing total-hits.