Forward Optional projections after final SSA

[1sgtpepper]

Summary

Part of #29343.

Removes redundant lowered Optional wrappers in the post-codegen peephole pass. The rewrites require compiler-generated identity, exact operands, and safe use counts; differing fallbacks retain both selections.

Test Plan

• compiler and execution coverage for zero, nonzero, input, and dynamic-key fallbacks
• negative coverage for lookalikes, mismatched conditions, and escaping values
• CircleCI

[1sgtpepper]

@mohammadfawaz I kept alias forwarding limited to generated Optionals because enabling it in the normal const-prop path changed 35 unrelated expectations. Would you keep it narrow here too?

[IGI-111]

I don't really like this solution, having partial application of some pass using some new syntax is confusing and error prone. We don't really know what has or hasn't been lowered at a given point.

I'd much rather we delay all options lowering and handle it in the intermediary passes if necessary or (and this is probably preferred) have a late optimization pass that notices and simplifies the pattern reliably.

[1sgtpepper]

Yeah, that makes sense. The partial const-prop entry point is awkward, and one
concrete issue is that it only
handles finalize contexts—the same shape still shows up in
views.
So I agree this should not stay as a special AST subphase.

The late peephole sounds like the better fit. I'd keep OptionLowering where it is,
carry generated-Optional provenance through codegen, and match the exact
ternary/cast/ternary sequence there. That keeps imported Optional__... structs and
unrelated casts out of the rule. Delaying OptionLowering looks considerably larger
because flattening and codegen already assume the lowered representation. Is that
roughly the direction you had in mind?