Currently, when handling a series of S3 API calls, keep-web builds a new filesystem object for each call. In some common workflows (reading several small files from a large collection / deep in a project tree), this will be very inefficient.

Initial implementation may be a map[apitoken]sitefs, using a single sitefs to serve many read-only requests until reaching a configured TTL or performing a write operation with the same token. This should be relatively simple, and although write operations would still be inefficient, sequences of read operations would be much faster.