Option to skip scratch

[wlandau]

When I run this drake benchmark on Windows, file.rename() is a bottleneck (second flame graph here). Would you be open to a global cache option to bypass scratch/ and write directly to the final destination? I know scratch/ prevents race conditions in the general case, but for RDS storrs, drake never tries to write to the same file from two concurrent processes.

Related: #114.

[wlandau]

Hmm... there is a chance this will conflict with #80 in some cases. But drake has changed a lot since then, including a couple changes that should avoid the need for atomic writes. I will investigate.