Make WandbLogger run initialization lazy for non-spawn distributed operation

[awaelchli]

What does this PR do?

Fixes #17487

Prior to 2.0, when strategy="ddp_spawn" was the default distributed strategy, the wandb logger had to solve the following problem:

• Avoid creating empty duplicate runs in main process and rank-0 worker process
• Avoid creating multiple runs when repeatedly calling fit/validate/test.

These were addressed by enabling the wandb-service feature (#11650), which would re-attach subprocesses to the one run created by the main process. This would avoid creating duplicated runs under different circumstances. However, the drawback to the solution is that the experiment had to be created in the main process for the service to reattach properly. This is inconsistent with other loggers who create their runs lazily, and leads to inconveniences like in #17487.

This PR removes the run initialization in WandbLogger.__init__ to make it truly lazy. The justification is that today in 2.0, the default strategy is ddp proper, with which the above issues never existed. To not break the usage with ddp_spawn, for users who still use that strategy, I "hacked" the __getstate__ so that when mp.spawn is called, the special wandb-service still gets enabled (and the run gets initialized in the main process).

cc @Borda @awaelchli @morganmcg1 @borisdayma @scottire @parambharat