[Perf] Reduce register-allocation spills on hot kernels

[Simonsbs]

Goal

Improve register allocation quality to reduce spills/reloads on hot paths.

Scope

• Add spill diagnostics per function (count and bytes moved).
• Tune linear-scan heuristics for live-range splitting and preferred register reuse.
• Add regression tests that assert spill reductions on selected kernels.

Acceptance

• Spill count metric reported in perf artifacts.
• At least 20% spill reduction on 3 hot kernels.
• Corresponding runtime improvement in those kernels.