[Feat] Nonautoregressive Methods

[cbhua]

Description

Contributed by @Furffico, @tycbony, and @henry-yeh from DeepACO and GLOP. Adding new support for non-autoregressive (NAR) methods, including edge embedding methods for various environments, anisotropic GNN encoder with edge-gating mechanism, non-autoregressive decoder, and related utils functions.

Motivation and Context

NAR methods are an important extension for RL4CO model support, aside from various previous autoregressive methods.

Types of changes

• New feature (mostly just adding new functionality)
• Example (update in the folder of examples)
• Breaking change (there are some changes to basic classes, even though most of them are simple code-level optimization, just be careful about hidden bugs)

Checklist

• This change requires an extension of the documentation.
• This change has provided a testing.

[cbhua]

Great code, nice work.

[Furffico]

I have modified the code following the review above. Please review the changes.

[cbhua]

Great work, thanks!