Better stable branch hints

[aumetra]

(This is a relatively small change, hence no prior discussion in an issue)

This PR adapts the technique outlined in hashbrown for branch hints on stable, utilizing the #[cold] attribute.
In my local benchmarks I observed an improvement of ~8-13% on average (but I'm not quite sure how many of these were flukes).

[Licenser]

This looks really cool :) thanks for sharing it and contributing it!

[Licenser]

The differences don't seem to be in the 8% range but they are measurable through the benchmarks :D well done 👍

group                                                    base                                   changes
-----                                                    ----                                   -------
apache_builds/simd_json::to_tape_with_buffers            1.01     90.7±4.85µs  1337.7 MB/sec    1.00     90.2±2.78µs  1345.7 MB/sec
canada/simd_json::to_tape_with_buffers                   1.00      3.6±0.14ms   601.4 MB/sec    1.00      3.6±0.18ms   599.0 MB/sec
citm_catalog/simd_json::to_tape_with_buffers             1.01  1115.1±81.18µs  1477.2 MB/sec    1.00  1109.5±82.41µs  1484.7 MB/sec
event_stacktrace_10kb/simd_json::to_tape_with_buffers    1.01      3.7±0.29µs     2.5 GB/sec    1.00      3.7±0.23µs     2.5 GB/sec
github_events/simd_json::to_tape_with_buffers            1.00     40.0±1.52µs  1553.9 MB/sec    1.00     39.9±2.04µs  1556.9 MB/sec
log/simd_json::to_tape_with_buffers                      1.00  1551.0±85.81ns  1337.4 MB/sec    1.00  1557.2±104.51ns  1332.0 MB/sec
twitter/simd_json::to_tape_with_buffers                  1.01   437.4±27.53µs  1376.8 MB/sec    1.00   434.6±28.27µs  1385.7 MB/sec
Warning: Failed to comment: HttpError: Resource not accessible by integration
Commenting is not possible from forks.
┌─────────┬─────────────────────────────────────────────────────────┬──────────────────┬──────────────────────┬────────────┐
│ (index) │                          name                           │   baseDuration   │   changesDuration    │ difference │
├─────────┼─────────────────────────────────────────────────────────┼──────────────────┼──────────────────────┼────────────┤
│    0    │     'apache_builds/simd_json::to_tape_with_buffers'     │  '90.7±4.85µs'   │  '**90.2±2.78µs**'   │  '-0.99'   │
│    1    │        'canada/simd_json::to_tape_with_buffers'         │   '3.6±0.14ms'   │     '3.6±0.18ms'     │   '0.0'    │
│    2    │     'citm_catalog/simd_json::to_tape_with_buffers'      │ '1115.1±81.18µs' │ '**1109.5±82.41µs**' │  '-0.99'   │
│    3    │ 'event_stacktrace_10kb/simd_json::to_tape_with_buffers' │   '3.7±0.29µs'   │   '**3.7±0.23µs**'   │  '-0.99'   │
│    4    │     'github_events/simd_json::to_tape_with_buffers'     │  '40.0±1.52µs'   │    '39.9±2.04µs'     │   '0.0'    │
│    5    │          'log/simd_json::to_tape_with_buffers'          │ '1551.0±85.81ns' │  '1557.2±104.51ns'   │   '0.0'    │
│    6    │        'twitter/simd_json::to_tape_with_buffers'        │ '437.4±27.53µs'  │ '**434.6±28.27µs**'  │  '-0.99'   │
│    7    │                           ''                            │    undefined     │      undefined       │   '+NaN'   │
└─────────┴─────────────────────────────────────────────────────────┴──────────────────┴──────────────────────┴────────────┘

[aumetra]

Heh, I guess the 8% were just my laptop acting up a bit then

[Licenser]

:D no harm it's an improvement one way or the other, doesn't always need to be 8%, this gives 1% in the tests, and might very well do more in other situations. What matters is that it's better in all situations so it's not a fluke but a through the board improvement 🚀