Cache generated code

[Kijewski]

• Speeds up compilation of a project a lot, when only one, a few, or no templates were changed.
• Speeds up compilation a little even for an initial compilation.
• Far better error messages in generated code.
• The generated code is easily debuggable.

[Kijewski]

I'll tidy up the PR after y'all tell me what you think, e.g. introducing the use of IndexMap, and deriving Hash in their own commits.

[Kijewski]

I could suppress the newly generated clippy warnings, too, but actually clippy is quite right with its warnings. Same condition in multiple elif-blocks, calculating X / X, etc. are 99% programming errors (outside of tests).

[djc]

Do you have benchmark numbers? In a bunch of other projects, I've adopted a different approach that I think might be better: generate code in an integration test. Here's one example, I've done a few others if you're interested:

mattsse/chromiumoxide#80

So the idea would be that we have an askama_build crate that contains the generator and other stuff, which offers a builder API that lets you pass a string like crate::foo::Bar with a Path to the template for that type, and this would generate the Template implementations into a file (formatted with rustfmt) written into the src dir as part of an integration test. This is nice because:

• No build script, build scripts slow down compilation pipelines and are a supply chain hazard
• No recompilation of the templates unless needed, no procedural macro expansion except at test time
• Integration test run in CI makes sure that the generated code is never out of date