Fix FP16 weight detection in NeuralNetworkBuilder.make_updatable (only first layer was checked)

[Anai-Guo]

Summary

NeuralNetworkBuilder.make_updatable() is supposed to refuse models that contain FP16 weight params, but _check_fp16_weight_param_exists() only ever looked at the first layer: its return False was indented inside the for layer in layers: loop.

So a model like [ip1 (FP32), ip2 (FP16)] was silently marked updatable.

This PR also fixes two more bugs in the same function:

• uniDirectionalLSTM: it did return self._check_fp16_weight_params_lstms(...), so the scan stopped at the first LSTM layer even when that LSTM was FP32.
• branch / loop: it read .float16Value from ifBranch / elseBranch / conditionNetwork / bodyNetwork. Those are nested NeuralNetwork messages with no such field, so this raised AttributeError. It now recurses into their .layers.

Why the test suite didn't catch it

MLModelUpdatableTest.test_nn_partial_fp16_make_updatable_fail was defined twice. The second definition replaced the first, so the first one never ran. That first one covers exactly this case: quantize every layer except ip1 to FP16, then expect make_updatable to raise.

The second definition tests something else (quantizing an already-updatable model), so I renamed it to test_nn_updatable_fp16_quantize_fail. Both tests run now.

Verification

Linux, coremltools==9.0 wheel (its builder.py is byte-identical to main), with test_model_updatable.py from this branch:

	pytest test_model_updatable.py
before (old builder.py)	1 failed, 16 passed: test_nn_partial_fp16_make_updatable_fail, AssertionError: ValueError not raised
after	17 passed

I also called _check_fp16_weight_param_exists directly on hand-built layer lists:

layers	before	after
[fp32 ip, fp16 ip]	False ❌	True
[fp32 LSTM, fp16 ip]	False ❌	True
[branch containing fp16 ip]	AttributeError ❌	True
[loop containing fp32 ip]	AttributeError ❌	False
[fp32 ip, loop containing fp32 ip]	False	False

🤖 Generated with Claude Code