Fix: failures during experimental feature parallel compile

[ac-mmi]

This PR addresses issue #6250, which highlights the need for clear execution separation between Tensors and SymbolicTensors.
It improves the apply() function in TensorFlow.js by explicitly handling synchronous (Tensor-based) and asynchronous (SymbolicTensor-based) execution separately.

1. Clear separation of sync and async execution:
1.1 If inputs are Tensors, execution happens synchronously.
1.2 If inputs are SymbolicTensors, execution happens asynchronously.

2. Ensured backward compatibility:
2.1 No change in how inputs are processed.
2.2 Layer building, weight loading, and input validation remain intact.

Currently, the apply() function mixes both execution types in a single flow, making it harder to debug and optimize. This PR makes it explicit when the function runs in sync or async mode, helping both performance and readability.

Fixes #6250

[google-cla]

Thanks for your pull request! It looks like this may be your first contribution to a Google open source project. Before we can look at your pull request, you'll need to sign a Contributor License Agreement (CLA).

View this failed invocation of the CLA check for more information.

For the most up to date status, view the checks section at the bottom of the pull request.