Be flexible on Content-Type when loading model weights/topology

[dsmilkov]

tensorflow/tfjs-core#1455 introduced an unnecessary strict Content-Type check that broke several users (and use-cases).

We should not be strict about the Content-Type and ONLY report it to the user IF we fail to parse the weight/topology in order to help them debug the issue.

Broken use-cases include:

• A server that was serving the weights with Content-Type text/html and the user is in no control of the content-type.
• Our Keras models on GCP (e.g. MobileNet 1.0 224) serve model.json as application/octet-stream, however we fail because we expect application/json.

[pyu10055]

@dsmilkov we can make the content type a warning message.
The model eventually get parsed by the either JSON or PB parser, and the weight data is parsed by the weightLoader, they will all generate some parse errors. (which are sometimes hard to understand, but not sure our code can interpret them either).

[nsthorat]

I vote for rolling the change back entirely and patching the rollback into the release branch. I think a warning would be annoying, especially for those without control of the server.

To fix the original issue, I think checking that the bytes line up / doing sanity checks on the parsed JSON is the most robust way to do this.