Bugfixes

• fix handling of build dependencies with # in path

Performance

• improve memory usage when using the filesystem cache

When reading a file into a Buffer and picking smaller slices of the Buffer
the small slices still keep a reference to the large original Buffer.
The leads to increased memory usage. A fix would be to clone the slice into
a smaller Buffer when wasting too much memory, but this has a performance cost.
There is now a new option cache.allowCollectingMemory which controls that.
For one-off builds you probably want allowCollectingMemory: false and
for watch builds you probably want allowCollectingMemory: true.
It defaults to false in production mode and true in development mode.