perf: speed up iter, range, scan methods with Rc<Node> (#356)

This PR improves iteration performance by storing nodes as Rc<Node<K>>.

The change is inspired by the Entry API PoC in #251, but applied
separately to avoid breaking changes.

btreeset:
```
Summary:
  instructions:
    status:   Regressions and improvements 🔴🟢
    counts:   [total 100 | regressed 7 | improved 40 | new 0 | unchanged 53]
    change:   [max +2.21K | p75 -179.95K | median -207.33K | p25 -620.79K | min -2.49M]
    change %: [max +5.01% | p75 -0.04% | median -1.11% | p25 -5.50% | min -11.48%]
```

btreemap:
```
Summary:
  instructions:
    status:   Regressions and improvements 🔴🟢
    counts:   [total 285 | regressed 3 | improved 8 | new 0 | unchanged 274]
    change:   [max +5.34M | p75 +411.31K | median +146.65K | p25 +97.32K | min -1.49M]
    change %: [max +3.36% | p75 +0.06% | median +0.03% | p25 +0.01% | min -10.92%]
```

Most of the improved benchmarks related to `iter`, `range`, `scan`
methods.
Some regressions are small and limited to benchmarks not involving
traversal.

UPD(from
[comment](#356 (comment))):
At first glance using `Rc<Node<K>>` should not help since we don't
explicitly clone `Node` anywhere.
The improvement comes from **avoiding repeated moves of large value**.
`Node` is fairly big (multiple fields and vectors) comparing to `Rc`, so
moving it around during iteration adds overhead.
Switching to `Rc` makes those moves much cheaper, it's just a pointer
copy.

Author: maksymar