Refactor, extend robots and machines lists, and add datacenter/browser access detection

[slint]

AI disclosure: did some iterations/exploration with Claude on the Classifier "builder" API to arrive at what is now committed. Also used some smaller Claude subagents to batch-classify and split some of the more generic UAs in machine_extra.txt from the original set dumped from the Zenodo web access logs.

• fix: machine takes precedence over robot
  The COUNTER robot list and the Make-Data-Count machine list overlap almost
  entirely (34 of the 35 machine patterns are also in the robot list), so wget,
  curl, python and friends matched both is_robot and is_machine. With filter_robots
  running before flag_machines downstream, that traffic was excluded as robots
  instead of counted as machine access.

  COUNTER CoP for Research Data 10.5 requires the robot blacklist to leave out the
  general-purpose agents researchers use and to stay a subset of the COUNTER list.
  is_robot now returns false when the machine list also matches, so a user agent in
  both lists is a machine. is_robot_or_machine is unchanged.

• refactor: composable, class-based Classifier API
  Restructure the flat module functions into a Classifier built by a fluent
  ClassifierBuilder, so lists can be extended without editing the package. Lists are
  added as presets, callables that add sources to the builder, applied with
  builder.use(preset); counter_preset is the generic COUNTER baseline. A deployment
  can add its own patterns with builder.robots(file_patterns(path), ignore_case=True).
  _PatternSet supports both case-sensitive and case-insensitive matching.

  The module-level is_robot / is_machine / is_robot_or_machine remain, backed by a
  default classifier built from counter_preset, so existing callers are unchanged.

• feat: extended preset from the crawler-user-agents dataset
  Add extended_preset, which builds on the COUNTER baseline using the maintained
  crawler-user-agents dataset (added as a dependency, so updates arrive through the
  package). Rather than its single mixed regex, the preset reads the raw entries and
  splits them by tag: http-library and browser-automation tools become machines,
  every other tag becomes a robot, matched case-sensitively as the dataset intends.
  This catches the modern crawlers (CamelCase bots, AI crawlers, link-preview bots)
  the frozen, case-sensitive baseline lists miss.

  Also add a small curated machine_extra.txt of non-browser tools and CLIs the
  dataset does not cover (geospatial, astronomy, bioinformatics and workflow tools,
  reference managers, download managers), matched case-insensitively. These were
  observed accessing research-data repositories and are reusable beyond any one
  instance; deployment-specific patterns still go through the builder.

• feat: datacenter/hosting ASN classification
  Add network-origin detection to the classifier, to catch automation that fakes a
  real browser user agent but runs from cloud/hosting infrastructure. is_datacenter
  classifies an ASN against datacenter_asn.txt (generated by scripts/update-asn-list.py
  from the union of brianhama/bad-asn-list and PeeringDB Content networks) minus an
  allow list (Apple iCloud Private Relay). is_browser exposes a conservative
  browser-UA heuristic. extended_preset also adds the datacenter ASN lists.

  is_datacenter_ip resolves an IP to its ASN through a resolver supplied to the
  builder; maxminddb_resolver (the asn extra) builds one over a GeoLite2-ASN mmdb with
  an in-memory cache. The library reads no config and ships no geo database; only ASN
  numbers are vendored.