optim: Store indices instead of making new SmartStrings

[cdown]

About ~12% of time is spent in doing conversions from &str to SmartString, which may require a reallocation anyway.

1. https://github.com/bodil/smartstring/blob/e407ca23c747257a812d2d2e70bf336412718c3a/src/lib.rs#L646
2. (or if inline): https://github.com/bodil/smartstring/blob/e407ca23c747257a812d2d2e70bf336412718c3a/src/inline.rs#L83

Maybe we can just store the indices instead of converting to and from SmartStrings? We only need a Borrowed anyway at self.get() time. No idea if it would actually be faster.

[cdown]

Oh, we do actually need key later as part of error reporting. Hmm.

I wonder if this in general would be better encapsulated with a single Formatter struct that holds format pieces (at first an empty Vec, populated on first run), formatters, and the format string. Then we could safely hold on to the &str because we still have the backing format string in this struct.