[SECURITY] `torch.load()` called without `weights_only=True` under `torch>=1.0` pin

Bug description

transformers4rec.torch.model.base.Model.load (rewritten in PR #808, commit 41b14d7b) and its accompanying test call torch.load(...) without passing weights_only=True:

• Test call site:
  https://github.com/NVIDIA-Merlin/Transformers4Rec/blob/8bf122f5dcb39feecfc6dabde734d79c2d1c4380/tests/unit/torch/model/test_model.py#L425
• Docstring guidance in Model.load directs users toward the same pattern.

requirements/pytorch.txt currently pins only torch>=1.0:

https://github.com/NVIDIA-Merlin/Transformers4Rec/blob/8bf122f5dcb39feecfc6dabde734d79c2d1c4380/requirements/pytorch.txt

On installs of PyTorch <2.6 (which the pin permits), torch.load defaults to weights_only=False. That is unrestricted pickle deserialization: any __reduce__ on the file executes on load, which is the same class of issue as CVE-2025-32434. PyTorch ≥2.6 flipped the default to weights_only=True, but T4Rec's pin does not enforce that.

This is particularly relevant because PRs #802/#804 were framed as pickle-security hardening — the hardening does not cover this code path. When the allowlist-based utils/serialization.py was removed in PR #808, torch.load became the primary load path again, with its version-dependent default.

Steps/Code to reproduce bug

Benign repro (side effect is a temp file; no persistent change):

import os, tempfile, torch

class Exploit:
    def __reduce__(self):
        return (os.system, ("touch /tmp/t4rec_rce_demo",))

with tempfile.NamedTemporaryFile(suffix=".pt", delete=False) as tf:
    torch.save(Exploit(), tf.name)
    path = tf.name

torch.load(path, weights_only=False)    # runs os.system under pickle's __reduce__
assert os.path.exists("/tmp/t4rec_rce_demo")

torch.load(path, weights_only=True) refuses the same payload with UnpicklingError:

Weights only load failed. [...] Trying to load unsupported GLOBAL posix.system whose module posix is blocked.

Expected behavior

• Every internal torch.load call passes weights_only=True.
• Tutorials and docstrings are updated to use the same pattern.
• requirements/pytorch.txt bumps the floor to torch>=2.6 so the safe default is always in effect.

Environment details

• Transformers4Rec: main @ 8bf122f5
• PyTorch: the pin is >=1.0 (see requirements/pytorch.txt); installs on <2.6 are the affected population.
• Python: any

Additional context

Two-part remediation:

1. In code: add weights_only=True to every torch.load(...) call in the library, tests, and example notebooks. This is safe because the surrounding code already expects a state_dict (plain tensor dict), which weights_only=True supports.

2. In requirements: bump torch>=1.0 to torch>=2.6. The project's other dependencies (HF transformers, Merlin) already require a recent PyTorch, so this bump is effectively a no-op for supported users.

Happy to send a PR covering both parts.

Start with transformers4rec.torch.model.base.Model.load, its docstring, and the call in tests/unit/torch/model/test_model.py around line 425. Search the library, tests, and example notebooks for other torch.load calls, then inspect requirements/pytorch.txt. Done means every load uses weights_only=True, guidance matches, and the relevant unit tests pass with the PyTorch floor set to 2.6.