Finalize repeat expr inference behaviour with inferred repeat counts

[BoxyUwU]

I believe this should be the last change of how repeat exprs are handled before it's finished for generic_arg_infer. Assuming we don't wind up deciding to replace this all with a new predicate kind :)

This PR has three actual changes:

• Always defer the checks to end of typeck even when generic arg infer is not enabled (needs an FCP)
• Properly handle element exprs that are constants when the repeat count is inferred
• "Isolate" each repeat expr check so that inference constraints from Copy goals dont affect other repeat expr checks resulting in weird order-dependent inference

The commit history and tests should be relatively helpful for understanding this PR's impl.

r? compiler-errors