Add Data.Aeson.Decoding working on token stream

[phadej]

There are two motivation points:

1. it's faster
2. it's more expressive

The first point is actually accidental, but nevertheless very nice to have. In benchmarks decoding strict bytestring through Tokens seems to be 30% or so faster for examples we have:

 Examples
decode
github-issues
lazy: OK (0.14s)
2.03 ms ± 168 μs
strict: OK (0.25s)
1.94 ms ± 180 μs
tokens: OK (0.34s)
1.27 ms ± 61 μs
Twitter
decode
Generic
direct
twitter100: OK (1.30s)
2.43 ms ± 161 μs
jp100: OK (0.25s)
2.09 ms ± 126 μs
twitter100t: OK (0.33s)
1.34 ms ± 121 μs
jp100t: OK (0.17s)
1.37 ms ± 86 μs
Manual
direct
twitter100: OK (0.21s)
1.76 ms ± 86 μs
jp100: OK (0.30s)
2.29 ms ± 207 μs
twitter100t: OK (0.74s)
1.38 ms ± 33 μs
jp100t: OK (0.20s)
1.61 ms ± 125 μs
TH
direct
twitter100: OK (0.27s)
2.08 ms ± 180 μs
jp100: OK (0.54s)
2.17 ms ± 149 μs
twitter100t: OK (0.36s)
1.38 ms ± 50 μs
jp100t: OK (0.39s)
1.48 ms ± 49 μs

The second point is also quite important.
If some users really need to know the order of key-value pairs in a JSON object, or care whether number was integral, decimal or scientific; now they can differentiate.

Though they cannot use FromJSON machinery, this is a start.

E.g. issues like Should we make a clear value-level distinction between floating and integral numbers?#546 (For Value it would be bad, for Tokens it's fine).

[phadej]

For now these additions would live only in Data.Aeson.Decoding. For next major we can replace ones in Data.Aeson with them.

[phadej]

The lazy bytestring token decoder is also fast, but slower than using strict bytestring as input.

All
Examples
decode
github-issues
lazy: OK (0.22s)
1.66 ms ± 114 μs
strict: OK (0.19s)
1.55 ms ± 96 μs
lazy tok: OK (0.32s)
1.23 ms ± 66 μs
strict tok: OK (0.13s)
1.02 ms ± 93 μs

[phadej]

I also run the benchmark with libperf: less instructions and seems to be half of cache references, which may explain the speedup.

All
Examples
decode
github-issues
lazy: OK (0.04s)
10.3 M instructions
2.17 M branch instructions
54.0 K branch misspredictions (2.5%)
569 K cache references
62.5 K cache misses (11.0%)
strict: OK
6.96 M instructions
1.53 M branch instructions
41.4 K branch misspredictions (2.7%)
327 K cache references
42.7 K cache misses (13.1%)
lazy tok: OK
6.59 M instructions
1.54 M branch instructions
42.9 K branch misspredictions (2.8%)
245 K cache references
30.5 K cache misses (12.4%)
strict tok: OK
4.71 M instructions
1.11 M branch instructions
26.6 K branch misspredictions (2.4%)
157 K cache references
15.2 K cache misses (9.6%)
Twitter
decode
Manual
direct
twitter100: OK
5.86 M instructions
1.16 M branch instructions
27.5 K branch misspredictions (2.4%)
213 K cache references
27.5 K cache misses (12.9%)
jp100: OK
7.88 M instructions
1.63 M branch instructions
39.0 K branch misspredictions (2.4%)
275 K cache references
42.0 K cache misses (15.3%)
twitter100t: OK
4.21 M instructions
930 K branch instructions
23.2 K branch misspredictions (2.5%)
105 K cache references
17.8 K cache misses (16.9%)
jp100t: OK
4.93 M instructions
1.11 M branch instructions
25.5 K branch misspredictions (2.3%)
104 K cache references
15.8 K cache misses (15.1%)