[NuclioRuntime] Always pass serving graph to pod using config-map

[alxtkr77]

https://iguazio.atlassian.net/browse/ML-6223

[gtopper]

Looks good, but I'm wondering if we can keep the flexibility so we're not married to a hard coded path under /tmp.

[theSaarco]

@alxtkr77 - looking at this, I actually think there is an issue here with BC. If the user is running a function that runs with an older mlrun client version that doesn't support passing the spec as CM (i.e. reading it from file), with this change there is no way to make it work except upgrading the client version.
Before making this change constant, most serving graphs would pass as-is in env variable, and the issue may happen only for large serving graphs (which arguable wouldn't work anyway due to their size). Now we're introducing a breaking change that will happen for every serving function deployed using an older client.
We need to give the user an option to keep using env variables for the serving spec using a configuration on the runtime or something similar. Trying to automatically deduce whether the client is old or not is problematic since we don't always know what version of mlrun is included in the image used.

[theSaarco]

LGTM - one small fix needed in the test.

[theSaarco]

You need to revert these changes at the end of the test. Since you have monkeypatch available, best to use it to perform these patching, as it will take care of reverting automatically.