GetTempPath hot in some benchmarks in Microsoft.ML.PerformanceTests

I have noticed that depending on the length of the temporary path, some benchmarks see a significant performance drop because of the need to convert from a Windows short name to a long name.

If I use the temporary directory path `C:\tempdir\`, the benchmark gives these performance results on my machine:

```
Microsoft.ML.PerformanceTests.exe --filter '*CreatePredictionEngine' --envVars TEMP:C:\tempdir TMP:C:\tempdir

| Method | Mean | Error | StdDev | Extra Metric |
|----------------------- |---------:|--------:|--------:|-------------:|
| CreatePredictionEngine | 249.5 us | 4.80 us | 4.93 us | - |
```

If I use a long temporary directory `C:\Users\username\AppData\Local\Temp\ML.NET\A Very Very Long File Name That Will Be Converted To A Short Path\Temp` (which has the short name `C:\Users\username\AppData\Local\Temp\ML.NET\AVERYV~1\Temp`), the benchmark shows significantly worse performance on the same machine:

```
.\Microsoft.ML.PerformanceTests.exe --filter '*CreatePredictionEngine' --envVars TEMP:C:\Users\username\AppData\Local\Temp\ML.NET\AVERYV~1\Temp TMP:C:\Users\username\AppData\Local\Temp\ML.NET\AVERYV~1\Temp

| Method | Mean | Error | StdDev | Extra Metric |
|----------------------- |---------:|---------:|---------:|-------------:|
| CreatePredictionEngine | 10.53 ms | 0.082 ms | 0.077 ms | - |
```

Looking at this under PerfView, I see a lot of extra time spent [here](https://github.com/dotnet/machinelearning/blob/f385b06aa0aeb64000fc341ecc31e43057c1ee13/src/Microsoft.ML.Core/Environment/HostEnvironmentBase.cs#L330).

I'm wondering if it is possible to cache this lookup? It seems misleading for the benchmark performance to fluctuate so much because of the temporary directory path.