fix(analysis): solve loop-aware buffer lifetimes

[zhen8838]

Why

• Fixes fix(analysis): give a loop-carried value the residency the loop gives it #169: loop-carried SSA values need physical reuse constraints, not summed logical lifetimes or unconditional carry exemptions.

What

• Record structured SSA intervals plus real/synthetic use events, so carry ties reject later old-value uses without a second HIR traversal.
• Traverse HIR nodes with AllocationConstraintVisitor; derive exact loop-aware operand-to-result relations from the existing polyhedral access maps.
• Compile independent insert_slice result/destination and result/update relations only inside each gmem/smem address space.
• Keep every logical SSA box in CP-SAT, seed one complete feasible assignment, and return after the first validated placement.
• Report rmem as the largest logical value and lock all 52 existing Analysis corpus cases to exact per-level peaks.

Contract

• gmem/smem peak_bytes is the returned placement high-water; rmem is the largest projected value.
• Capacity remains a post-solve, non-fatal MemoryMetadata.errors/JSON/# error report.
• solver_status=feasible does not claim a minimum.
• New liveness/allocation helpers remain package-internal.

Risk

• Scheme B preserves iteration-dependent offsets in ISL but compiles translated subviews as containment. Exact affine byte offsets remain a localized future upgrade in apply_constraints().