[core] Introduce MemoryPoolConfigurator class hierarchy

[hnyls2002]

Summary

• Introduce MemoryPoolConfigurator base class with unified coeff+bias interface (calculate_pool_sizes / calculate_pool_sizes_from_max_tokens)
• Add DefaultPoolConfigurator for MHA/MLA/NSA/FP4 — absorbs get_cell_size_per_token, num_layers deduction, DFLASH scaling
• Add HybridSWAPoolConfigurator for Gemma2/Command-R/MiMo — absorbs resolve_hybrid_swa_tokens with full/swa pool splitting
• Add create_memory_pool_configurator() factory
• Rewrite _resolve_memory_pool_config to use configurator flow
• Delete absorbed methods: profile_max_num_token, _resolve_hybrid_swa_tokens
• Move MemoryPoolConfig from model_runner_kv_cache_mixin.py to pool_configurator.py
• Page alignment now owned by configurator; removed from _apply_token_constraints

Follows up on #22384. Mamba configurator is a separate follow-up.

Behavioral changes

• Fix hybrid SWA _cell_size to use ratio-weighted formula (F*nf + r*S*ns), so --max-total-tokens correctly constrains full_tokens rather than inflating it through a memory-budget round-trip (pre-existing issue in the old code)
• Configurator returns MemoryPoolConfig directly; max_running_requests default changed from required int to Optional[int] = None (filled by consumer after configurator runs)

Test plan

• /rerun-stage stage-a-test-1
• /rerun-stage stage-b-test-small-1-gpu
• /rerun-stage stage-b-test-large-1-gpu

[gemini-code-assist]

Warning

You have reached your daily quota limit. Please wait up to 24 hours and I will start processing your requests again!

[hnyls2002]

/rerun-test test_swa_unittest.py test_mimo_models.py test_deepseek_v3_mtp.py test_dsa_models_mtp.py test_qwen3_next_models_mtp.py test_qwen35_models.py test_triton_sliding_window.py test_mamba_unittest.py test_mamba2_mixer.py test_nvidia_nemotron_nano_v2.py test_nvidia_nemotron_3_super_bf16.py test_mla_deepseek_v3.py test_generation_models.py

[github-actions]

✅ 1-gpu-h100 (4 tests): View workflow run

cd test/ && python3 registered/unit/mem_cache/test_swa_unittest.py
cd test/ && python3 registered/attention/test_triton_sliding_window.py
cd test/ && python3 registered/mla/test_mla_deepseek_v3.py
cd test/ && python3 registered/models/test_generation_models.py

✅ 8-gpu-h200 (3 tests): View workflow run

cd test/ && python3 registered/8-gpu-models/test_mimo_models.py
cd test/ && python3 registered/8-gpu-models/test_dsa_models_mtp.py
cd test/ && python3 registered/8-gpu-models/test_nvidia_nemotron_3_super_bf16.py

✅ 4-gpu-h100 (1 test): View workflow run

cd test/ && python3 registered/4-gpu-models/test_qwen3_next_models_mtp.py

✅ 4-gpu-b200 (1 test): View workflow run

cd test/ && python3 registered/4-gpu-models/test_qwen35_models.py

✅ 1-gpu-5090 (1 test): View workflow run

cd test/ && python3 registered/unit/mem_cache/test_mamba_unittest.py

✅ 2-gpu-h100 (2 tests): View workflow run

cd test/ && python3 registered/layers/mamba/test_mamba2_mixer.py
cd test/ && python3 registered/models/test_nvidia_nemotron_nano_v2.py

❌ test_deepseek_v3_mtp.py: Ambiguous filename test_deepseek_v3_mtp.py — matched 2 files:

• test/registered/8-gpu-models/test_deepseek_v3_mtp.py
• test/registered/amd/test_deepseek_v3_mtp.py

Please provide the full path, e.g. /rerun-test test/registered/8-gpu-models/test_deepseek_v3_mtp.py