Performance investigation

Splitting out the benchmark discussion from https://github.com/JuliaMath/FixedPointDecimals.jl/pull/36.

------

Here's the benchmark in a gist:
https://gist.github.com/NHDaly/a8fae0d1d65ab1066c585c27e54146fa

And the results in a google spreadsheet:
https://docs.google.com/spreadsheets/d/1Lc3ughgwwK25cpwbnuLRxw19EtaMxdCsMmtnspT_4H8/edit?usp=sharing

-------

Here are the results:

  |   | Operation | Values |   |   |   |   |   |   |   |  
-- | -- | -- | -- | -- | -- | -- | -- | -- | -- | -- | --
  |   | identity |   | ÷ |   | + |   | / |   | * |  
Category | Type | time (ms) | allocs | time (ms) | allocs | time (ms) | allocs | time (ms) | allocs | time (ms) | allocs
Int | Int32 | 1.35 | 0 | 5.16 | 0 | 1.47 | 0 | 2.35 | 0 | 1.61 | 0
  | Int64 | 1.86 | 0 | 18.66 | 0 | 1.89 | 0 | 2.46 | 0 | 1.95 | 0
  | Int128 | 3.77 | 0 | 26.07 | 0 | 3.85 | 0 | 16.74 | 0 | 3.95 | 0
Float | Float32 | 1.35 | 0 | 28.97 | 0 | 1.47 | 0 | 1.75 | 0 | 1.47 | 0
  | Float64 | 1.85 | 0 | 27.37 | 0 | 1.88 | 0 | 2.45 | 0 | 1.89 | 0
FixedDecimal | FD{Int32,2} | 1.35 | 0 | 5.16 | 0 | 1.48 | 0 | 38.20 | 0 | 31.73 | 0
  | FD{Int64,2} | 1.86 | 0 | 18.75 | 0 | 1.89 | 0 | 59.18 | 0 | 47.03 | 0
  | FD{Int128,2} | 1320.01 | 14000000 | 26.19 | 0 | 1324.35 | 14000000 | 7267.32 | 72879639 | 6139.13 | 62000000

--------------

Here are my current question:

- [x] For some reason, even just element-wise copying an array of `FixedDecimal{Int128, 2}` into another array, allocates like crazy. (14,000,000 allocations / 1,000,000 elements).
- Where do these allocations come from?
- [ ] `/` and `*` of `FixedDecimal{Int64, 2}` are more expensive than for `FixedDecimal{Int32, 2}`., by a factor of around 1.5x each, whereas `/` and `*` for `Int64` and `Int32` are almost identical.
- My current _guess_ is that this might be related to promoting to `Int128` during those operations (due to `widemul`), which seems to be slower than `Int64` across the board.
- [ ] `/` for `Int128` is like 6x slower than for `Int32`! Where does that come from?