fix(irodori-tts): do not cast projector input to the weight dtype

nn.quantize turns the projection into a QuantizedLinear whose weight is
packed uint32, so aligning the backbone activations to
self.projector.weight.dtype truncated them to integers. This was a no-op
in fp16, so only quantized v4 builds were affected: the text/caption
projector output landed 79% off and generation was audibly broken.

Dropping the cast brings the 8-bit projector back to 0.7% of fp16, and
text_state to 0.7% of the fp32 PyTorch reference. Adds a regression test
that quantizes a small v4 model and checks the projector output scale.

Co-Authored-By: Claude Opus 5 <noreply@anthropic.com>