fix: [Memory Estimator] Fix text config attribute fetching for new multimodal models (e.g., Qwen-3-VL)

[LZY2275]

Background
Most modern multimodal models extended from text models (e.g., Qwen-3-VL) encapsulate core text-related attributes (vocab_size, max_position_embeddings, hidden_size, etc.) in hf_config.text_config. However, some older multimodal models (e.g., Qwen-2-vl-7B-Instruct) keep these attributes at the top level of the config.

The original code only fetched attributes from the top-level config, which worked for older models like Qwen-2-vl but failed for newer ones like Qwen-3-VL (since their text attributes are nested in text_config).

Changes Made
Add support for text_config-nested attributes:

• Fetch vocab_size and max_position_embeddings from either top-level config or text_config sub-object

[ISEEKYAN]

thank you @LZY2275 for your interests in the memory estimator, I did not tested this on qwen-vl models and no vision part is developed. I think now only the language part can be precisely estimated, do you have any estimation results on qwen-vl? Do you think it necessary to precisely estimate the vision part? Feel free to discuss with me!