[Refactor] Complete TSFit Removal - Enable Clean Multi-Backend Architecture

[astrogilda]

Is your feature request related to a problem? Please describe.
TSFit represents significant technical debt that violates every core engineering principle in our standards:

• Zero Tolerance for Technical Debt: TSFit IS technical debt - 400+ lines of duplicate code
• Single Responsibility Principle: Handles 6+ responsibilities in one monolithic class
• Performance: Creates 7.66x overhead compared to direct backend usage
• Architecture: Blocks clean multi-backend design, limiting users to ~5 models instead of 30+

Current issues:

• Duplicate implementations between TSFit and backend system
• TSFitBestLag reimplements what StatsForecast's AutoARIMA already provides
• Monolithic design prevents leveraging each backend's strengths
• Forces all models through same interface, losing backend-specific optimizations

Describe the solution you'd like
Complete removal of TSFit in favor of the already-implemented multi-backend architecture:

1. Phase 1: Add missing features to backends (AR support, HQIC, data rescaling)
2. Phase 2: Add deprecation warnings to TSFit
3. Phase 3: Migrate ~250 tests from TSFit to backends
4. Phase 4: Remove TSFit implementation files
5. Phase 5: Clean up all use_backend parameters and conditionals

End state - Clean multi-backend architecture:

├── StatsForecastBackend (30+ models, Auto*, batch optimization)
├── StatsModelsBackend (VAR/VECM, econometrics, diagnostics)
└── ArchBackend (GARCH family, volatility modeling)

Describe alternatives you've considered

• Keep TSFit as legacy compatibility layer - violates Zero Tolerance for Technical Debt
• Gradual deprecation over multiple versions - prolongs the architectural debt
• Refactor TSFit to use backends internally - adds unnecessary abstraction layer

Additional context

• TSFit is NOT used in production bootstrap code (only in tests)
• Backend system already primary (use_backend=True by default)
• Bootstrap classes use ModelFittingService, not TSFit directly
• ~2,000 lines of code can be removed
• Unlocks StatsForecast's full model suite (AutoETS, Theta, MSTL, etc.)

Benefits:

• Performance: Maintain 7.66x speedup already achieved
• Architecture: Enable specialized backends for different use cases
• Maintainability: Single source of truth for each functionality
• Future-proofing: Clear path for MLForecast/NeuralForecast integration