Reduce memory consumption from attribute accumulation

[e-n-f]

The externally-visible effect is that -zg maxzoom guessing now takes the number of features with duplicate locations into account. The main calculation is still the same, and only takes into account distinct locations, but the base zoom is then increased if necessary to ensure that the number of features present in each zoom level, distinctly visible or not, is reasonable.

The internal change is that attribute names in serial_feature are now stored as shared pointers, so attribute names can be shared between features instead of having to be stored individually.

[ChrisLoer]

Notes from review:

Generally looks good -- we're trading a certain drop in memory usage from the two optimizations against a few new uncertainties:

• Maybe tiling takes a lot longer from going to higher zooms in some cases
• Maybe low zooms are changed too much in those datasets with lots of duplicates from having chosen to go to higher zooms?
• Maybe running time of tiling is longer because of extra indirection with shared_ptr optimization, plus costs of looking up keys in the map?

But all of those unknowns have at least some coverage in our existing tests, so we're OK with the fact that there might be some places where they're a net loss.

But before merging, let's add at least one automated test case for this duplicate -> higher zoom pathway.

[ChrisLoer]

Thanks for the test!