From the last commit, we had this for the big SVG file in #574:

--------------------------------------------------------------------------------
  n        time(i)         total(B)   useful-heap(B) extra-heap(B)    stacks(B)
--------------------------------------------------------------------------------
 30 22,796,106,012    1,553,581,072    1,329,943,324   223,637,748            0
                                                       ^^^^^^^^^^^

That extra-heap is a lot of allocator metadata or waste space,
possibly from heap fragmentation due to all the realloc() shenanigans
from using into_boxed_slice() on the starting Vec<PathCommand>.

This commit makes PathBuilder use a SmallVec:

    pub struct PathBuilder {
        path_commands: SmallVec<[PathCommand; 32]>,
    }

That is, it will keep up to 32 PathCommand directly inside itself, and
only spill to a heap allocation if more elements come in.

We still into_boxed_slice() for the final Path.  Hopefully this makes
realloc() find heap blocks that are more closely packed.  Indeed:

--------------------------------------------------------------------------------
  n        time(i)         total(B)   useful-heap(B) extra-heap(B)    stacks(B)
--------------------------------------------------------------------------------
 33 24,139,598,653    1,416,831,176    1,329,943,212    86,887,964            0
                                                        ^^^^^^^^^^

That's a lot of less waste space.  Note also how the total bytes
shrinks from 1,553,581,072 to 1,416,831,176.