fix(B614): suppress false positive on non-literal weights_only; add torch.jit.load regression coverage

[DevamShah]

Summary

Improves the B614 ("unsafe PyTorch load") check on two fronts, both reducing false positives:

1. Suppress B614 when weights_only is a non-literal expression (a variable, attribute, subscript, call, etc.). When the value cannot be resolved statically, the check now gives the caller the benefit of the doubt rather than emitting a finding — mirroring the existing behaviour of B615 (huggingface_unsafe_download), which uses the same weights_only / non-literal handling.
2. Regression coverage that torch.jit.load is not flagged (refs B614 False Positive when using torch.jit.load #1293). torch.jit.load uses TorchScript serialization, not pickle, and is correctly outside B614's matched set (torch.load / torch.serialization.load). The example file now documents this so the behaviour reported in B614 False Positive when using torch.jit.load #1293 is locked in by a test fixture.

Problem / motivation

• Non-literal weights_only (the substantive fix): today torch.load(f, weights_only=flag) is flagged because the check only treats the literal True / "True" as safe. Real code frequently passes weights_only as a computed value (config flag, function arg), and flagging those is a false positive. B615 already solved the identical problem the same way; this brings B614 into line.
• torch.jit.load (B614 False Positive when using torch.jit.load #1293): users reported B614 firing on torch.jit.load. On current main its qualified name (torch.jit.load) is not in B614's matched set, so it is already not flagged — this PR adds explicit example coverage so that stays true.

Change

• bandit/plugins/pytorch_load.py: when the matched call (torch.load / torch.serialization.load) passes weights_only as a non-ast.Constant value, return without a finding. The existing weights_only == True/"True" safe-path and the genuine-unsafe path are unchanged. Uses the same context._context.get("call") keyword-inspection idiom as the merged B615 plugin.
• examples/pytorch_load.py: adds non-triggering cases — torch.jit.load (with and without map_location) and a non-literal weights_only — alongside the existing genuinely-unsafe calls.

Trade-off (intentional, precedented)

Suppressing on a non-literal weights_only means torch.load(f, weights_only=flag) where flag is False at runtime will not be flagged (a false negative). This matches B615's deliberate "benefit of the doubt" stance for unresolvable arguments — favouring precision over recall for low-confidence static cases. The version-aware behaviour also requested in #1293 (treating weights_only as defaulting to True on torch ≥ 2.6) is intentionally out of scope here and left for a follow-up, since it requires version detection.

Testing / validation

• tests/functional/test_functional.py::FunctionalTests::test_pytorch_load passes with the unchanged expectation (MEDIUM: 3, HIGH: 3) — the added cases correctly yield zero new findings.
• Direct run on examples/pytorch_load.py: B614 fires only on the three genuinely-unsafe torch.load calls; it does not fire on torch.jit.load or the non-literal weights_only case.
• Adjacent functional tests pass; no regressions.