Add qfloat8 offloading

[PlagueKind]

Add qfloat8 (quanto) support with layer offloading and LoRA/LoKr merge

Quanto's qfloat8 quantization was forced to float8 (torchao) in several places because the memory offload and merge-in code paths only recognized torchao's tensor subclass. This adds proper detection for quanto tensors and removes the resulting workarounds.

Changes

• toolkit/memory_management/manager_modules.py — _is_ao_quantized_tensor now also recognizes optimum.quanto.* tensor subclasses, not just torchao.*. Fixes offload/pin/bounce for qfloat8 weights, which were previously silently treated as plain float tensors.
• toolkit/config_modules.py — removed the forced qfloat8 → float8 override for layer_offloading, now unnecessary.
• toolkit/util/quantize.py — is_quantized_tensor now recognizes quanto tensors; requantize_module_weight now re-quantizes quanto modules via .freeze() when no torchao config is available (quanto modules retain their own weight_qtype/optimizer config from creation).
• toolkit/network_mixins.py, toolkit/models/lokr.py, toolkit/lora_special.py — removed early-returns that silently skipped merge_in/merge_out entirely for quanto-quantized weights ('weight._data' in state_dict). The dequantize → merge → re-quantize logic underneath already worked generically once detection was fixed.
• extensions_built_in/diffusion_models/krea2/krea2.py — removed the forced qtype override for assistant LoRA loading, now unnecessary since merge_in correctly re-quantizes.

Testing

• Verified _is_ao_quantized_tensor / _move_params_to_cpu_and_pin against a real quanto QLinear: outer wrapper and inner _data/_scale storage both correctly land on CPU, not just the outer tensor.
• Verified steady-state VRAM under offload is bounded prefetch-ring cost (PIPELINE_DEPTH × weight size), not per-layer accumulation.
• Ran toolkit/memory_management/test_memory_manager.py (temporarily swapped to qfloat8) across all 8 offload/checkpoint configs on a 1.2B-param model with a real LoRA network — no exceptions, correct output/gradient match against non-offloaded reference (output diff ~2e-3, input grad diff ~1e-6, consistent with fp8 quant noise).
• Verified merge_in/merge_out round-trip correctness for both FullModule and LoRAModule against real quanto QBytesTensor weights — weight changes as expected, stays quantized throughout, round-trip diff ~4e-3 (quant noise only).