[6.0] Support closures that capture opened pack element types

[slavapestov]

Partial 6.0 cherry-pick of #73712 and previous PRs.

• Description: Closures can capture values of pack element type, which are introduced by a repeat expression or a for ... in repeat loop. For example, here we accidentally capture a pack element type, because && takes an autoclosure:

  protocol P { var foo: Bool { get } }
  
  func f<each T: P>(_ ts: repeat each T) {
    for t in repeat each ts {
      _ = true && t.foo
    }
  }
  

  This would cause crashes and type checking errors. This PR implements the SIL lowering support for closures that capture pack elements. Some solver support is incomplete, however a bunch of cases now work.

• Scope of the issue: Affects many users, and is even easier to hit now with for ... in repeat loops.

• Origination: This never worked.

• Risk: I believe Low for code that doesn't use packs, Medium for code using packs. There's a bunch of stuff here:

  • I redid ASTVerifier support for local archetypes. This is off in an noassert build.
  • I changed the HasArchetype bit in TypeBase into HasPrimaryArchetype and added a hasPrimaryArchetype() predicate to test it, but the existing hasArchetype() predicate has the same behavior as before. I only updated a few callers to use the new hasPrimaryArchetype().
  • There are some new assertions. I made a workaround in autodiff to simulate the old noassert behavior in a narrow case
  • Capture analysis now tracks uses of local archetypes. Any existing code that didn't crash before should end up getting an empty list of local archetype captures here. But a logic error might break existing code that uses packs.
  • Type lowering now consults the list of captured archetypes. If it's not empty, a bunch of new code runs to build a new generic signature, extend the substitution map for the call, etc. I hooked this into the existing mechanism which drops the generic signature for a closure in generic context that has no generic captures. It should all short-circuit trivially if the list of captured archetypes is empty.
  • There is a new pass that runs after SILGen to rewrite SIL instructions in a non-trivial way. This pass only runs if there were captured local archetypes.
  • There's a representation change for variables declared within the scope of a for ... in repeat loop. Some extra book-keeping that used to exist is now subsumed into my new logic. This might break existing code, but pack iteration is a new feature in 6.0. This change was separately reviewed on main by @simanerush.
  • For the most part validation tests pass at each intermediate commit so it should be easy to bisect any regressions.

• Reviewed by: review pending - @hborla @rjmccall @simanerush

• Radar: rdar://113505724, rdar://122293832, rdar://124329076.

[slavapestov]

@swift-ci Please test

[slavapestov]

@swift-ci Please test source compatibility

[eeckstein]

I reviewed 28fd20c and f6d6774#diff-9bcf6c807c755531489d2fbe8aeaa54308636dd1879f26a6d7917f08f6a5fb55R130-R152: LGTM

[slavapestov]

(swift-distributed-actors fails with the same error in the baseline source compat run)

[slavapestov]

@swift-ci Please test