Revert to float32 — ONNX Runtime Web WASM does not support float16 tensors

Float16 models produce Cast nodes that ORT Web WASM cannot execute.
Reverting to the original float32 weights (297 MB total).

An int8 quantization path exists for the encoder (34 MB, 75% reduction)
but the decoder cannot be quantized due to the If-node wrapping from
optimum's merged model export. Future work: extract subgraphs and
quantize, or re-export decoder without If merge.

decoder_model_merged_fp16.onnx → decoder_model_merged.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:b4c4edd133304675f8c4b6e4044890b27dbd9f4012929aea602f7a6307ad6766
-size 84467797
+oid sha256:8d9f918d4c6fe763cd234bc3ad0124debb678767760a6877fbb1ea66c30286f2
+size 168092405

encoder_model_fp16.onnx → encoder_model.onnx

@@ -1,3 +1,3 @@
 version https://git-lfs.github.com/spec/v1
-oid sha256:0abb24ec99bc0e0921523e3b31800f019100fe239f9acdba8a349156a806a128
-size 71268793
+oid sha256:991fe56183b8506ea158bdf8c49df49cb1ca5c125770f42adf1f2401bd065e5f
+size 142347416