This can have a huge overall impact on mypy performance when search paths are long

This is a pretty small win, it's below the noise floor on
macrobenchmark, but if you time the hashing specifically it saves about
100ms (0.5%) on `python -m mypy -c 'import torch' --no-incremental`.
blake2b is slower

See #17948

There's one call site which has varargs that I leave as os.path.join, it
doesn't show up on my profile. I do see the `endswith` on the profile,
we could try `path[-1] == '/'` instead (could save a few dozen
milliseconds)

In my work environment, this is about a 10% speedup:
```
λ hyperfine -w 1 -M 3 '/tmp/mypy_primer/timer_mypy_6eddd3ab1/venv/bin/mypy  -c "import torch" --no-incremental --python-executable /opt/oai/bin/python'
Benchmark 1: /tmp/mypy_primer/timer_mypy_6eddd3ab1/venv/bin/mypy  -c "import torch" --no-incremental --python-executable /opt/oai/bin/python
  Time (mean ± σ):     30.842 s ±  0.119 s    [User: 26.383 s, System: 4.396 s]
  Range (min … max):   30.706 s … 30.927 s    3 runs
```
Compared to:
```
λ hyperfine -w 1 -M 3 '/tmp/mypy_primer/timer_mypy_88ae62b4a/venv/bin/mypy  -c "import torch" --no-incremental --python-executable /opt/oai/bin/python'
Benchmark 1: /tmp/mypy_primer/timer_mypy_88ae62b4a/venv/bin/mypy  -c "import torch" --no-incremental --python-executable /opt/oai/bin/python
  Time (mean ± σ):     34.161 s ±  0.163 s    [User: 29.818 s, System: 4.289 s]
  Range (min … max):   34.013 s … 34.336 s    3 runs
```

In the toy "long" environment mentioned in the issue, this is about a 7%
speedup:
```
λ hyperfine -w 1 -M 3 '/tmp/mypy_primer/timer_mypy_6eddd3ab1/venv/bin/mypy  -c "import torch" --no-incremental --python-executable long/bin/python'
Benchmark 1: /tmp/mypy_primer/timer_mypy_6eddd3ab1/venv/bin/mypy  -c "import torch" --no-incremental --python-executable long/bin/python
  Time (mean ± σ):     23.177 s ±  0.317 s    [User: 20.265 s, System: 2.873 s]
  Range (min … max):   22.815 s … 23.407 s    3 runs
```
Compared to:
```
λ hyperfine -w 1 -M 3 '/tmp/mypy_primer/timer_mypy_88ae62b4a/venv/bin/mypy -c "import torch" --python-executable=long/bin/python --no-incremental'
Benchmark 1: /tmp/mypy_primer/timer_mypy_88ae62b4a/venv/bin/mypy -c "import torch" --python-executable=long/bin/python --no-incremental
  Time (mean ± σ):     24.838 s ±  0.237 s    [User: 22.038 s, System: 2.750 s]
  Range (min … max):   24.598 s … 25.073 s    3 runs
```

In the "clean" environment, this is a 1% speedup, but below the noise
floor.

See #17948

This is about 1.06x faster on `mypy -c 'import torch'` (in both the
clean and openai environments)
- 19.094 -> 17.896 
- 34.161 -> 32.214

```
λ hyperfine -w 1 -M 3 '/tmp/mypy_primer/timer_mypy_36738b392/venv/bin/mypy  -c "import torch" --no-incremental --python-executable clean/bin/python'
Benchmark 1: /tmp/mypy_primer/timer_mypy_36738b392/venv/bin/mypy  -c "import torch" --no-incremental --python-executable clean/bin/python
  Time (mean ± σ):     17.896 s ±  0.130 s    [User: 16.472 s, System: 1.408 s]
  Range (min … max):   17.757 s … 18.014 s    3 runs

 λ hyperfine -w 1 -M 3 '/tmp/mypy_primer/timer_mypy_36738b392/venv/bin/mypy  -c "import torch" --no-incremental --python-executable /opt/oai/bin/python' 
Benchmark 1: /tmp/mypy_primer/timer_mypy_36738b392/venv/bin/mypy  -c "import torch" --no-incremental --python-executable /opt/oai/bin/python
  Time (mean ± σ):     32.214 s ±  0.106 s    [User: 29.468 s, System: 2.722 s]
  Range (min … max):   32.098 s … 32.305 s    3 runs
```

For `mypy -c 'import torch'`, the cache load time goes from 0.44s to
0.25s as measured by manager's data_json_load_time. If I time dump times
specifically, I see a saving of 0.65s to 0.07s. Overall, a pretty
reasonable perf win -- should we make it a required dependency?

See also #3456

See #17948
This is starting to show up on profiles

- 1.01x faster on clean (below noise)
- 1.02x faster on long
- 1.02x faster on openai
- 1.01x faster on openai incremental

I had a dumb bug that was preventing the optimisation for a while, I'll
see if I can make it even faster. Currently it's a small improvement

We could also get rid of the legacy stuff in mypy 2.0

See #17948

- 1.01x faster on clean
- 1.06x faster on long
- 1.04x faster on openai
- 1.26x faster on openai incremental

Follow up to #17955