feat(inference): add layer-aligned KV cache management (#1964)

[mrveiss]

Summary

• LayerKVCache: per-layer (k, v) tensor pair storage with lazy allocation
• KVCacheManager: memory estimation, max sequence length calculator, RTX 4070 preset
• KVCacheConfig: validated config with dtype mapping (fp16/bf16/fp32 + aliases)
• Lazy torch import — module loads without GPU/torch installed
• 50 tests across 12 test classes

Closes #1964

Test plan

• Config validation (dtype bytes, aliases, 6 error paths)
• Cache get/update: single/multi-layer, multi-step accumulation
• Shape validation: wrong batch/heads/head_dim/rank
• Trim: reduce, to-zero, no-op, negative, content correctness
• Memory estimation: formula, fp32 vs fp16, scaling
• Max seq_len: positive, monotonic, tiny budget floor, round-trip
• RTX 4070 helper: positive and within budget
• flake8 passes (F821 suppressed for torch.Tensor type hints)

🤖 Generated with Claude Code

[github-actions]

✅ SSOT Configuration Compliance: Passing

🎉 No hardcoded values detected that have SSOT config equivalents!