[Aikido] Fix security issue in transformers via minor version upgrade from 5.1.0 to 5.14.0

[aikido-autofix]

Upgrade transformers to fix critical RCE vulnerabilities in model loading that bypass trust_remote_code, plus path traversal and arbitrary file read issues in tokenizer and config deserialization.

✅ 4 CVEs resolved by this upgrade, including 1 critical 🚨 CVE

This PR will resolve the following CVEs:

Issue	Severity	Description
CVE-2026-5241	🚨 CRITICAL	[transformers] A vulnerability in LightGlue model loading allows attackers to execute arbitrary code by overriding the trust_remote_code=False parameter with untrusted configuration data, enabling remote code execution despite explicit security settings. This affects API servers, notebooks, and CI/CD pipelines, risking credential theft and system compromise.
CVE-2026-4372	HIGH	[transformers] A malicious config.json file can execute arbitrary Python code when loading a model via AutoModelForCausalLM.from_pretrained(), bypassing the trust_remote_code security mechanism. This remote code execution vulnerability exploits unfiltered deserialization of configuration attributes during standard model loading.
AIKIDO-2026-295187	MEDIUM	[transformers] A tokenizer loading vulnerability allows arbitrary local files to be read when loading from untrusted repositories via crafted paths in tokenizer_config.json, enabling information disclosure through save/push workflows even with trust_remote_code disabled.
AIKIDO-2026-614493	MEDIUM	[transformers] Path traversal vulnerability in model loading allows crafted local models to read .safetensors files outside the model directory via untrusted transformers_weights values in config.json, enabling unauthorized weight access and path disclosure.