Unexpected behavior of _TempDirWithCustomCleanup

## Describe the bug
This is not 100% a bug in `datasets`, but behavior that surprised me and I think this could be made more robust on the `datasets`side.

When using `datasets.disable_caching()`, cache files are written to a temporary directory. This directory should be based on the environment variable TMPDIR. I want to set TMPDIR at runtime using os.ENVIRON["TMPDIR"] = something, but depending on other imported modules this can fail to take effect.

## Steps to reproduce the bug
`_TempDirWithCustomCleanup` relies on `tempfile` to generate a path to a temporary directory. However, `tempfile` generates the path only once. This can be a problem when trying to set TMPDIR at runtime whenever other code imports `tempfile` first and does something unexpected.

For example (after too much trial and error) I found out that a different part of the code base I work with defines a class `PatchedDataCollatorForLanguageModeling(transformers.DataCollatorForLanguageModeling)` based on a `transformers` class. This import is enough to trigger `tempfile` to generate `tempfile` to generate a temporary path and leading to the wrong path being cached in `tempfile.tempdir`.

## Suggestion:
I could file this also as bug with `transformers`, but I think fixing this on the datasets would be much more robust:
Datasets could recompute the temporary path once (technically possible via `tempfile._get_default_tempdir` or resetting
the global variable `tempfile.tmpdir` to None) before setting its own global `_TEMP_DIR_FOR_TEMP_CACHE_FILES`.