Fix torchao_convert, remove StretchedAffineQuantizedTensor

[lisjin]

Follow-up to QuantOptimizer.torchao_convert added in #2947:

1. Use the "_default" key for the dict passed to ModuleFqnToConfig instead of listing all linear weight names.
2. Fix unwanted quantization of tied weights when using "_default". Some HF models have a _tied_weights_keys variable (e.g., in Qwen3Model).
   i. Suppose we tie lm_head.weight with model.embed_tokens.weight. Even though they're tied, their names both show up in model.named_parameters(). This is a problem since HF's quantization_config will try to apply the "_default" config to the tied LM head, which may differ from the embedding's. @metascroy This is a major issue for our mixed quantization of embeddings and LM head.
3. @jerryzh168 Remove StretchedAffineQuantizedTensor to align with deprecated version=1 in torchao configs.

[pytorch-bot]

🔗 Helpful Links

🧪 See artifacts and rendered test results at hud.pytorch.org/pr/pytorch/ao/3015

• 📄 Preview Python docs built from this PR

Note: Links to docs will display an error until the docs builds have been completed.

✅ No Failures

As of commit 3516435 with merge base 9a770a5 ():
💚 Looks good so far! There are no failures yet. 💚

This comment was automatically generated by Dr. CI and updates every 15 minutes.

[metascroy]

This is a problem since HF's quantization_config will try to apply the "_default" config to the tied LM head, which may differ from the embedding's

Can you say more here with an example code snippet?

[lisjin]

This is a problem since HF's quantization_config will try to apply the "_default" config to the tied LM head, which may differ from the embedding's

Can you say more here with an example code snippet?

@metascroy Sure, here is where HF applies quantize_ given a "_default" key. Their filter_fn doesn't doesn't check for tied parameters, so it would apply the default config to LM head tied to the embeddings.