Prefer SafeString to mark_safe where possible

[Tim McCurrach]

mark_safe takes roughly twice as long as simply creating a SafeString - using pyperf:

mark_safe: Mean +- std dev: 296 ns +- 12 ns
SafeString: Mean +- std dev: 158 ns +- 7 ns

There are many places in the django codebase where we know the thing we are marking as safe to be a normal (not marked as safe) string. In such cases it makes sense to use SafeString instead of mark_safe.

To play devils advocate, you could definitely argue that this is an unnecessary micro-optimisation. Following a brief search for mark_safe, there are some situations where we have something like mark_safe(X) and where evaluating X will take sufficiently long that any savings made marking the string as safe would be rendered insignificant.

Having said that, there are other places where we end up calling mark_safe a very large number of times. In such situations a small saving in time will add up to a larger saving. There are also places where we even have mark_safe("some string literal").

Furthermore, since this change is literally just replacing one word with an equally clear word, it would have no effect on complexity or readability, and so for a slight performance boost, why not?