Fix bugs and dead code across repo (#13762)

* Fix bugs and dead code across repo

Repo-wide bug scan: each change is a real bug fix or dead code removal,
no new abstractions, no defensive validation for impossible states.

Bugs:
- utils/general.py: Windows logger lambda captured fn by reference (late
  binding) so LOGGER.info ended up calling LOGGER.warning. Use default arg.
- utils/augmentations.py: LetterBox auto-mode `hs, ws = genexp if ... else
  self.h, self.w` parsed as `(genexp), self.w`, leaving hs as a generator.
  Parenthesize the tuple branches.
- utils/loggers/comet/__init__.py: `raise "<string>"` is a TypeError in
  Python 3 — promote to ValueError.
- models/common.py: TensorRT dynamic-shape path called `data.resize_()`
  on output bindings without refreshing binding_addrs, so execute_v2 ran
  against stale pointers if the tensor reallocated. Refresh after resize_.
- export.py: TFLite export leaked a file handle (open().write()).
- export.py: opset run() default (12) didn't match CLI default (17).
- benchmarks.py: `if map` referenced the Python builtin (always truthy);
  the alternate branch was unreachable. Drop the dead conditional.
- detect.py: CSV label ternary `names[c] if hide_conf else f"{names[c]}"`
  produced the same string in both branches.

Dead code / cleanup:
- val.py: commented-out duplicate argsort.
- export.py: redundant `weights_dir = None` overwritten on the next line.
- utils/loggers/comet/__init__.py: unreachable bare `return`.
- utils/segment/plots.py: commented-out matplotlib code referencing a
  non-existent API.

Perf:
- classify/predict.py: `torch.Tensor(im).to(device)` allocates on CPU then
  copies; switch to `torch.from_numpy(im).to(device)` to skip the CPU copy.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

* Auto-format by https://ultralytics.com/actions

* Revert two changes that caused CI regressions

- classify/predict.py: torch.Tensor(im) accepts both numpy arrays and
  Tensors; the dataset already calls ToTensor() so im is a Tensor by the
  time it reaches this line. torch.from_numpy rejects Tensors with
  'TypeError: expected np.ndarray'. Restore torch.Tensor.
- export.py: restore run() opset default to 12 to match prior behavior;
  the CLI default of 17 was already correct on master.

Co-Authored-By: Claude Opus 4.7 (1M context) <noreply@anthropic.com>

---------

Co-authored-by: Claude Opus 4.7 (1M context) <noreply@anthropic.com>
Co-authored-by: UltralyticsAssistant <web@ultralytics.com>