[quantization] [draft] GPTQ for VLM

[stamalakhov]

This PR is the first try out for full quantization of VLM model by GPTQ+PTQ.

TODO:

1. m.b. make it less resource intensive (right now it makes inference for the whole model, not in layerwise fashion)
2. support PTQ quantization
3. synchronize GPTQ/PTQ Conv3d quantization
4. support convert to circle

model	orig_accuracy_vqav2	minmax_quantize_accuracy_vqav2	GPTQ_mse_accuracy_vqav2	GPTQ_smse_accuracy_vqav2
Qwen2_2B	0.8900	0.8260	0.8450	0.8740
Qwen3_2B	0.8570	0.7970	0.8470	0.8390
Qwen3_4B	0.8950	0.8450	0.8910	0.8820

some_details

all models above were quantized using GPTQ+mse/GPTQ+smse:

1. weights of torch.nn.Linear, torch.nn.Conv2D, torch.nn.Conv1D, nn.ConvTranspose2d to 4 bits,
2. activations were left at float32
3. accuracy was computed on the first 1000 samples on vqav2

for 256:

model	vqav2_on_1000_samples
Qwen2_2B_original	0.8900
Qwen2_2B_GPTQ_mse_256_qsamples	0.8630
Qwen2_2B_GPTQ_smse_256_qsamples	0.8780
Qwen3_2B_original	0.8570
Qwen3_2B_GPTQ_mse_256_qsamples	0.8430
Qwen3_2B_GPTQ_smse_256_qsamples	0.8520

• Support Conv3d ([quantization] Support torch.nn.Conv3d in GPTQ #577)
• Support all Convs in SensitivityCalibrator ([quantization] Suppport convolutions in SensitivityCalibrator #581)
• Add example script ([quantization] Fallback for complex models #583)

Related: #548

TICO-DCO-1.0-Signed-off-by: s.malakhov s.malakhov@partner.samsung.com

[stamalakhov]

I believe we can close this draft. Everything was merged.