This avoids adding extra keys into the cache whenever these options change.

- Avoid actually compiling a module when it has same hash as an existing one
- Simplified the get/release lock stuff in some functions
- Added ability to log messages directly when calling _rmtree

Fix a crash following the new meaning of CudaNdarray.base to always point to the variable that contain the data.

This is needed to avoid keeping in memory a chain of intermediate
arrays.

There was no memory leak.

new function in strutil for making C code look nice even though it was generated programmatically with the wrong indentation