Fix 0.6B fine-tuning crash: project text embedding before adding to c…

[kru2710shna]

CAUSE
In finetuning/sft_12hz.py, the text and codec embeddings are added element-wise:

input_embeddings = input_text_embedding + input_codec_embedding

But they are defined at different widths in the talker model (modeling_qwen3_tts.py):

self.codec_embedding = nn.Embedding(config.vocab_size, config.hidden_size)
self.text_embedding  = nn.Embedding(config.text_vocab_size, config.text_hidden_size)

So text_embedding has width text_hidden_size and codec_embedding has width hidden_size. These differ for 0.6B Base (text_hidden_size=2048, hidden_size=1024), so the add fails with:

RuntimeError: The size of tensor a (2048) must match the size of tensor b (1024) at non-singleton dimension 2

FIX
The talker already defines text_projection (Qwen3TTSTalkerResizeMLP, mapping text_hidden_size to hidden_size) and applies it on the inference path before combining text and codec embeddings. The fine-tuning script omitted it. This PR projects the text embedding through text_projection before the add, matching inference.
Because text_projection uses bias=True, the mask is re-applied after projection so padded positions remain zero:

input_text_embedding = model.talker.model.text_embedding(input_text_ids)
input_text_embedding = model.talker.text_projection(input_text_embedding)   # text_hidden_size -> hidden_size
input_text_embedding = input_text_embedding * text_embedding_mask           # re-zero padding after projection

Verification
Reproduced the crash on Qwen3-TTS-12Hz-0.6B-Base with training data prepared via prepare_data.py:

Before: RuntimeError: tensor a (2048) must match tensor b (1024) at the embedding add.
After: training proceeds and the step completes with finite loss (Epoch 0 | Step 0 | Loss: 11.82).

Fixes #335