Optimize the parsing of unquoted attributes

[stof]

Instead of handling each char one by one, we read them in batch until a char for which we need a special handling (chars ending the attribute value, references and chars needing to throw an error).

I ran the loading benchmark on a modified fixture file in which I unquoted all class attributes containing a single class (this kind of replacement was the easiest to do).

Before: 116.11662387848
After: 114.08836126328

There is no difference on the unmodified benchmark, because it uses quoted attributes everywhere.

[stof]

note that this strategy of reading chars until the special ones is already used when reading quoted attribute values.

[goetas]

because of recent CS changes, this requires to be rebased

[stof]

@goetas done

[goetas]

Thanks