Optimise locking when storing types per key

[systemed]

In #877 @pa5cal identified a slowdown caused by locking when we write to attributeMap for a given layer, which we use to record the types for each attribute in the vector tile (eventually written into the vector_layers JSON).

We had some discussion of fixes in the thread, and then in true "you wait three months for a bus" fashion, we also have #912 which also provides a fix. This PR takes from those to provide the simplest possible solution:

• We keep a record of what keys per layer have already been written in thread_local storage
• We use an unordered_set, which means we can use count to avoid the find/end threading issue
• We aren't worried about subsequently checking the type: first one always wins

I think this should achieve the issue while keeping the code compact and readable for what is a fairly marginal feature (the vector_layers JSON isn't used by most consumers). I've done some moderate testing on this but would be happy to see further reports.