fix(optimizer): prefer cheaper variant when it covers overflow at lower absolute cost

[biranofer]

Summary

• costAwareScaleUp greedily picked the most cost-efficient variant (lowest cost/perReplicaCapacity) but overpaid when a cheaper-by-absolute-cost variant could cover the required overflow at lower total spend
• Add a per-iteration fallback check: if covering the overflow with the cheapest variant costs strictly less than one additional replica of the current variant, round down and let the loop assign the remainder to the cheaper variant
• Guard added: fallback only applies when the cheapest variant has headroom below its maxReplicas cap

Details

The N-1 terms cancel algebraically, so the check reduces to:

fallbackReplicas * cheapest.Cost < vc.Cost

On equality the more efficient variant is kept (strict <) — at equal cost it provides more capacity per dollar.

Fixes #1251

Test plan

• Existing unit tests pass (make test — all 113 specs green including TestScaleToZero/should respect maxReplicas during scale-up)
• Empirically validated on pokprod001: two-variant Llama-3.1-8B (primary TP=2 cost=10, v2 TP=1 cost=5), rate=15 req/s, 25-min run — primary held at peak=2 replicas, v2 absorbed all remaining demand up to max=10, 98.6% SLO

🤖 Generated with Claude Code

[biranofer]

I'm considering to add a CS (Cost Save) configuration parameter [0..1] that prefer the cheapest only if it saves CS of the cost, i,e.
if fallbackReplicas <= cheapestAvailable && float64(fallbackReplicas)*cheapestVC.Cost <(1-CS)*vc.Cost { replicasNeeded-- }

A reasonable default can be 0.2

[deanlorenz]

• please add unit tests.
• please resolve conflicts after rebase to main.

NIT

• cheapestVariantCapacity does not break ties. order may change. may be harder to debug and test.
• if current is capped then cheapest must cover all remaining -- unlikely to be cheaper than one current replica.

[ev-shindin]

@biranofer please rebase and address @deanlorenz review