Update command line interface - #3456

The old deeplabcut/cli.py and root dlc.py were unmaintained DLC2-era
tooling: not wired to the installed `dlc` console script (which only
launches the GUI), untested, and train_network/evaluate_network/
analyze_videos bypassed engine selection by importing
pose_estimation_tensorflow directly instead of the public API.

deeplabcut/cli.py is rewritten as a Typer app whose commands are thin
wrappers around deeplabcut.* (the same functions `import deeplabcut as
dlc` exposes), so engine routing (PyTorch vs. TensorFlow) is handled by
the API layer for free. Required parameters are explicit CLI arguments;
optional ones default to an UNSET sentinel and are only forwarded if the
user actually set them, so the CLI never redeclares (and risks drifting
from) the API's own defaults. Every optional parameter, whether or not it
has a named flag, is also reachable via a repeatable `--set key=value`
escape hatch (YAML-typed); supplying the same parameter both ways is a
hard error. Each command's --help lists its --set-only parameters via an
epilog computed from the real function signature, and --set-help prints
the delegate's full docstring, so neither can go stale.

deeplabcut/__main__.py now dispatches: bare `dlc` still launches the GUI
unchanged, while `dlc <command>` (including `dlc --help`) routes to the
new CLI. Root dlc.py becomes a deprecated shim forwarding to the same CLI.