Remove backwards symlinks

[maherthomsi]

Description of changes:
Make the NVIDIA /usr/lib/nvidia/tesla backwards-compatibility symlinks opt-in so EKS k8s 1.37+ nvidia variants can drop them, while k8s 1.36 and below keep them. The FHS library-path normalization previously created these legacy symlinks for every variant; we don't want to carry that forward on new variants, but must not change behavior for existing ones.

This changes the defaults to clean (no symlinks) and ships -compat templates that downstream variants render to restore the old behavior.

• nvidia-container-toolkit: generate-cdi-specs.service no longer passes --additional-symlinks /usr/lib/nvidia/tesla by default. Its ExecStart can now be overridden by a rendered drop-in, and a new generate-cdi-specs-exec-start-conf-compat template restores the flag. The unit is ordered After=settings-applier.service so a rendered drop-in is present before the oneshot runs.
• nvidia-k8s-device-plugin: the default exec-start template drops --cdi-enabled-hooks create-lib-symlinks; a new nvidia-k8s-device-plugin-exec-start-conf-compat template retains it.

Testing done:
Built an aws-k8s-1.36-nvidia x86_64 AMI consuming this kit and booted it as a node on a live EKS 1.36 cluster (nvidia.com/gpu allocatable = 1). Verified both paths:

• Clean defaults (build with no downstream overrides): device-plugin exec-start.conf had no create-lib-symlinks; /etc/cdi/nvidia.json contained 0 nvidia/tesla references - confirming the intended 1.37+ behavior.
• Compat templates (build with downstream overrides applied): device-plugin exec-start.conf = --cdi-enabled-hooks create-lib-symlinks; generate-cdi-specs ran with --additional-symlinks; /etc/cdi/nvidia.json contained 30 nvidia/tesla references; nvidia-smi worked (driver 580.178.04); host /usr/lib/nvidia/tesla present (49 symlinks).
• Confirmed generate-cdi-specs.service keeps RefuseManualStart=true and that a daemon-reload (not a restart) applies the drop-in before the oneshot runs.

Terms of contribution:

By submitting this pull request, I agree that this contribution is dual-licensed under the terms of both the Apache License, version 2.0, and the MIT license.