Shell_benchmarks: fixes minor issues

This MR fixes small issues related with the IO benchmarks:

• Shell_benchmarks: optimizes Io_stats.load_tree
  • Sorted folding of Irmin context is very bad for performance.
• Shell_benchmarks: Io_stats.load_tree now returns the full paths
  • Benchmarks expect the full paths but load_tree returned relative paths.
• Shell_benchmarks: fixes Key_map.to_seq
  • Bug fix. Key_map.to_seq was not completely seq-ized and therefore got OOM-killed when traversing a huge directory of Tezos mainnet context.
• Benchmark: skips GC for Generator.Calculated
  • Gc.compact is not required for Calculated, since the measurements are already obtained in it.

Not an issue but the following will be required for new IO benchmarks:a

• Benchmark: exposed Measure.Time, adding measure_lwt
  • Required to measure times of IO benchmarks