feat: Workload Mix Optimizer (M115)

[hlin99]

Summary

Add Workload Mix Optimizer (M115) — given benchmark data for multiple workloads (different models or request patterns), find the minimum total GPU instances while meeting per-workload SLA constraints.

Changes

• WorkloadMixOptimizer class in workload_mix.py with brute-force enumeration and budget pruning
• WorkloadSpec, WorkloadAllocation, MixOptimizationResult Pydantic models
• SLA compliance check using measured percentiles from benchmark data
• CLI workload-mix subcommand with --workload (repeatable), --total-gpus, Rich table + JSON output
• Programmatic optimize_workload_mix() API
• 32 new tests covering models, optimizer, convenience API, public imports
• Updated ROADMAP.md (M114 ✅, M115 🔄) and docs/iterations/current.md

Closes #253

[hlin99-Review-Bot]

✅ Approved by hlin99-Review-Bot

Idea Value: Strong addition — workload mix optimization is a natural next step after multi-backend comparison. Aligns well with the project's data-driven GPU planning goals.

Code Quality:

• Clean WorkloadMixOptimizer with brute-force + budget pruning — appropriate for the problem size
• Pydantic models well-defined with proper validators (WorkloadSpec, WorkloadAllocation, MixOptimizationResult)
• SLA check uses measured percentiles from benchmark data (no guessing ✓)
• CLI follows established patterns with Rich table + JSON output
• 32 tests covering models, optimizer, convenience API, and public imports
• docs/iterations/current.md updated ✓
• ROADMAP.md updated ✓

CI: All checks pass (lint, test 3.10/3.11/3.12)

LGTM 👍

[hlin99-Review-BotX]

✅ Approved by hlin99-Review-BotX

Idea Value: Solid — workload mix optimization is the logical extension for multi-workload GPU cluster planning. Aligns with project goals.

Code Quality:

• WorkloadMixOptimizer brute-force + pruning approach is clean and appropriate
• Pydantic models with proper validators (WorkloadSpec.name strip, weight > 0)
• _check_sla uses measured percentiles — data-driven ✓
• _is_better comparison logic is correct (SLA-met first, then fewer instances, then lower waste)
• CLI follows repo patterns (Rich table + JSON, argparse registration)
• 32 tests with good coverage across models, optimizer, convenience API, imports
• docs/iterations/current.md and ROADMAP.md updated ✓

CI: All checks pass (lint, test 3.10/3.11/3.12)

LGTM 👍 — 2nd approval, should auto-merge.